Method and apparatus for optimizing cost-based heuristic instruction scheduling

• US 5,367,687 A
• Filed: 07/07/1993
• Issued: 11/22/1994
• Est. Priority Date: 03/11/1991
• Status: Expired due to Term

• Current Assignee: Sun Microsystems Incorporated (Oracle Corporation)
• Original Assignee: Sun Microsystems Incorporated (Oracle Corporation)
• Inventors: Tarsy, Gregory, Woodard, Michael J.
• Primary Examiner(s): Shaw, Gareth D.
• Assistant Examiner(s): Toplu, Lucien
• Application Number:
  US08/088,418
• Time in Patent Office: 503 Days
• Field of Search: 364/DIG. 1, 364/DIG. 2
• US Class Current: 717/149
• CPC Class Codes: G06F 11/3428 Benchmarking
• CPC Class Codes: G06F 8/445 Exploiting fine grain parallelism, i.e. parallelism at instruction level run-time instruction scheduling G06F9/3836

Abstract

A method and apparatus for optimizing cost-based heuristic instruction scheduling for a pipelined processor is disclosed which has particular application to compile time instruction scheduling after code generation. Instruction scheduling is optimized by determining the optimal weights to be used by an apparatus for cost based heuristic instruction scheduling for a particular pipelined processor. The optimal weights are determined based on the lowest of the lowest costs incurred by different collections of interrelated weight sets. Each collection of interrelated weight sets comprises a randomly generated initial weight set and subsequent interrelated weight sets generated in a predetermined manner. The predetermined manner for generating subsequent weight sets facilitates rapid identification of the optimal weight set for a collection, and thereby rapid identification of the overall optimal weight set for the collections.

Claims

1. In a computer system comprising a pipelined processor for executing instructions of programs in a parallel and overlapping manner, and a compiler for compiling and generating said instructions, wherein said compiler has a scheduler for scheduling said instructions for execution on said pipelined processor, and said scheduler schedules said instructions using N weighted cost based heuristics, a method for empirically selecting a set of N weights for said scheduler to weigh said N cost based heuristics, said method comprising the steps of:

a) generating arbitrarily an initial trial set of N weights, initializing said scheduler with said arbitrarily generated initial trial set of N weights, generating a plurality of benchmark programs, compiling said benchmark programs using said compiler, accumulating scheduling costs determined by said scheduler for said benchmark programs, initializing a lowest accumulated scheduling cost of said benchmark programs to said accumulated scheduling cost, and selecting said initial trial set of N weights as the selected set of N weights for said scheduler;

b) generating sequentially a first plurality of additional trial sets of N weights in a first manner, one trial set of N weights at a time, each of said first plurality of additional trial sets of N weights being generated by systematically varying the immediately preceding trial set of N weights along an orthogonal dimension of a weight space formed by the N weights, one orthogonal dimension at a time, reinitializing said scheduler with each of said additional trial sets of N weights after each of their generations, regenerating said plurality of benchmark programs after each of said reinitializations, recompiling said regenerated benchmark programs using said compiler after each of said regenerations, reaccumulating scheduling costs determined by said scheduler for said benchmark programs during each of said recompilations, comparing each of said reaccumulated scheduling cost with said lowest accumulated scheduling cost to determine whether a new lowest accumulated scheduling cost is found after each of said reaccumulations, terminating said generation of additional trial sets of N weights in said first manner as soon as a new lowest accumulated scheduling cost is found, updating said lowest accumulated scheduling cost to equal the newly found lowest accumulated scheduling cost if a new lowest accumulated scheduling cost is found, and selecting the trial set of N weights that yields the new lowest accumulated scheduling cost over the previously selected set of N weights as the selected set of N weights for said scheduler if a new lowest accumulated scheduling cost is found; and

c) generating sequentially a second plurality of additional trial sets of N weights in a second manner, one additional trial set of N weights at a time, each of said additional trial sets of N weights being generated under said second manner by systematically varying the immediately preceding trial set of N weights along the last orthogonal dimension with the last systematic variation made under said first manner, reinitializing said scheduler with each of said additional trial sets of N weights after each of their generations, regenerating said plurality of benchmark programs after each of said reinitializations, recompiling said regenerated benchmark programs using said compiler after each of said regenerations, reaccumulating scheduling costs determined by said scheduler for said benchmark programs during each of said recompilations, comparing each of said reaccumulated scheduling cost with said lowest accumulated scheduling cost to determine whether a new lowest accumulated scheduling cost is found after each of said reaccumulations, terminating said generation of additional trial sets in said second manner as soon as no new lowest accumulated scheduling cost is found, updating said lowest accumulated scheduling cost to equal the last newly determined lowest accumulated scheduling cost if at least one newly determined lowest accumulated scheduling cost is found, and selecting the last trial set of N weights that yields the last new lowest accumulated scheduling cost over the previously selected set of N weights as the selected set of N weights for said scheduler if at least one newly determined lowest accumulated scheduling cost is found.

2. The method as set forth in claim 1, wherein,

said first manner of generating said first plurality of additional trial set of N weights in said step b) has a predetermined finite number of variations:

said step b) is also terminated when all of said predetermined finite number of variations for generating additional trial sets of N weights under said first manner have been made; and

said step c) is not performed when said step b) is terminated under said all variations have been made condition.

3. The method as set forth in claim 1, wherein, said method further comprises the steps of:

d) initializing a lowest of the lowest accumulated scheduling cost to equal the lowest accumulated scheduling cost and selecting the selected set of N weights as an initial ultimate selected set of N weights after performing said steps a) through c) once;

e) repeating said steps a) through c) a plurality of times, one repetition at a time, comparing the new lowest accumulated scheduling cost to said lowest of the lowest scheduling cost to determine whether a new lowest of the lowest accumulated scheduling cost is found after each of said repetitions of said steps a) through c), updating said new lowest of the lowest accumulated scheduling cost with the new lowest of the lowest accumulated scheduling cost after each of said repetitions if a new lowest of the lowest accumulated scheduling cost is found, selecting the selected set of N weights that yields the new lowest of the lowest accumulated scheduling cost at the end of each of said repetitions over the previous ultimate selected set of N weights as the ultimate selected set of N weights for said scheduler if a new lowest of the lowest accumulated scheduling cost is found, and terminating said repetition of said steps a) through c) when no new lowest of the lowest accumulated scheduling cost is found for a predetermined number of consecutive repetitions.

4. The method as set forth in claim 1, wherein said initial trial set of N weights comprises N randomly generated weights.

5. The method as set forth in claim 2, wherein said systematic varying of the immediately preceding trial set of N weights along an orthogonal dimension of a weight space formed by said weights, one orthogonal dimension at a time, comprises varying the weights of the immediately preceding trial set of N weights, one weight at a time, each weight being varied at least one time in a predetermined manner.

6. The method as set forth in claim 5, wherein each of said weight variations under said first manner of trial weight generation is performed by adding a variation value to the weight being varied.

7. The method as set forth in claim 6, wherein each of said variation values under said first manner of trial weight generation is computed by multiplying a current value of a first variable, a current value of a second variable and a current value of a third variable,

said first variable being varied from -1 to 1,

said second variable being varied incrementally in a predetermined manner from a first initial value over a first finite amount in increasing arithmetic increments, and

said third variable being varied incrementally in a predetermined manner from a second initial value over a second finite amount in fixed geometric increments.

8. The method as set forth in claim 7, wherein said first variable is varied from 1 to 6 in increasing arithmetic increments of 2 and 3.

9. The method as set forth in claim 7, wherein said third variable is varied from 1 to 1000 in fixed geometric increments of 10s.

10. The method as set forth in claim 6 wherein said systematic varying of the immediately preceding trial set of N weights along the last orthogonal dimension with the last systematic variation made under said first manner comprises repeatedly adding the last variation value of said first manner to the last weight varied in said first manner of each immediately preceding trial set of N weights.

11. In a computer system comprising a pipelined processor for executing instructions of programs in a parallel and overlapping manner, and a compiler for compiling and generating said instructions, wherein said compiler has a scheduler for scheduling said instructions for execution on said pipelined processor, and said scheduler schedules said instructions using N weighted cost based heuristics, an apparatus for empirically selecting a set of N weights for said scheduler to weigh said N cost based heuristics, said apparatus comprising:

a) first trial weight generation means for generating arbitrarily an initial trial set of N weights;

(b) second trial weight generation means coupled to said first trial weight generation means for generating sequentially a first plurality of additional trial sets of N weights by systematically varying the immediately preceding trial set of N weights in a first manner, one trial set of N weights at a time, each of said first plurality of additional trial sets of N weights being generated by systematically varying the immediately preceding trial set of N weights along an orthogonal dimension of a weight space formed by the N weights, one orthogonal dimension at a time;

(c) third trial weight generation means coupled to said second trial weight generation means for generating sequentially a second plurality of additional trial sets of N weights by systematically varying the immediately preceding trial set of N weights in a second manner, each of said second plurality of additional trial sets of N weights being generated by systematic varying the immediately preceding trial set of N weights along the last orthogonal dimension with the last systematic variation made under said first manner;

d) initialization means coupled to said first, second, and third trial weight generation means and said scheduler for initializing said scheduler with said arbitrarily generated initial trial set of N weights after its generation, and reinitializing said scheduler with each of said first and second plurality of additional trial sets of N weights generated under said first and second manners after each of their generations;

e) benchmark generation means coupled to said initialization means and said compiler for generating an identical collection of benchmark programs for compilation by said compiler after said initialization, and each of said reinitializations of said scheduler;

f) cost accumulation means coupled to said compiler for accumulating scheduling costs determined by said scheduler for said benchmark programs during each of said compilations of said collection of benchmark programs;

g) first update means coupled to said first, second, and third trial weight generation means, and cost accumulation means for initially updating a lowest accumulated scheduling cost to equal the accumulated scheduling cost of said benchmark programs after the compilation in response to said initial trial set of N weights, updating said lowest accumulated scheduling cost to equal a newly found lowest accumulated scheduling cost if a new lowest accumulated scheduling cost is found through the compilations in response to said first plurality of additional trial set of N weights, and updating said lowest accumulated scheduling cost with the last newly found lowest accumulated scheduling cost if at least one new lowest accumulated scheduling cost is found through the compilations in response to said second plurality of additional trial set of N weights;

h) first comparison means coupled to said accumulation means and said first update means for comparing the accumulated scheduling cost of said benchmark programs with said lowest accumulated scheduling cost to determine if a new lowest accumulated scheduling cost is found after each of said compilations in response to said first plurality of additional trial set of N weights, and comparing the accumulated scheduling cost of said benchmark programs with said lowest accumulated scheduling cost to determine if no new lowest accumulated scheduling cost is found after each of said compilations in response to said second plurality of additional trial set of N weights;

i) first termination means coupled to said second and third trial weight generation means and said first comparison means for terminating said generation of said trial sets of N weights under said first manner by said second trial weight generation means as soon as a new lowest accumulated scheduling cost is found, and terminating said generation of said trial sets of N weights under said second manner by said third trial weight generation means as soon as no new lowest accumulated scheduling cost is found; and

j) first selection means coupled to said first, second, and third trial weight generation means, and said first comparison means for selecting said initial trial set of N weights as the selected set of N weights, selecting the trial set of N weights yielding a new lowest accumulated scheduling cost over the previously selected set of N weights as the selected set of N weights while said trial sets of N weights are generated under said first manner, and selecting the last trial set of N weights yield a new lowest accumulated scheduling cost over the previously selected set of N weights as the selected set of N weights while said trial sets of N weights are generated under said second manner.

12. The apparatus as set forth in claim 11, wherein,

said second trial weight generation means has a predetermined finite number of variations for generating said trial sets of N weights under said first manner;

said termination means also terminates said generation of said first plurality of additional trial sets of N weights under said first manner by said second trial weight generation means when said second trial weight generation means has made all of said predetermined finite number of variations.

13. The apparatus as set forth in claim 11, wherein, said apparatus further comprises:

k) repetition means coupled to said first, second, and third trial weight generation means, said initialization means, said benchmark generation means, said cost accumulation means, said first update means, said first comparison means, said first termination means, and said first selection means for repeating a plurality of times, one repetition at a time, said functions performed by said first, second, and third trial weight generation means, said initialization means, said benchmark generation means, said cost accumulation means, said first update means, said first comparison means, said first termination means, and said first selection means;

l) second update means coupled to said repetition means and said first update means for updating a lowest of the lowest accumulated scheduling cost to equal said lowest accumulated scheduling cost after said functions of said first, second, and third trial weight generation means, said initialization means, said benchmark generation means, said cost accumulation means, said first update means, said first comparison means, said first termination means, and said first selection means are performed once, and updating said lowest of the lowest accumulated scheduling cost to equal said lowest accumulated scheduling cost after each of said repetitions if a new lowest of the lowest accumulated scheduling cost is found;

m) second comparison means coupled to said repetition means and said first and second update means for comparing said lowest accumulated scheduling cost with said lowest of the lowest accumulated scheduling cost to determine if a new lowest of the lowest accumulated scheduling cost is found after each of said repetitions;

n) second selection means coupled to said repetition means and said first selection means for selecting the selected set of N weights as an initial ultimate selected set of N weights after said functions of said first, second, and third trial weight generation means, said initialization means, said benchmark generation means, said cost accumulation means, said first update means, said first comparison means, said first termination means, and said first selection means are performed once, and selecting the selected set of N weights over the previous ultimate selected set of N weights as the ultimate selected set of N weights after each of said repetitions if a new lowest of the lowest accumulated scheduling cost is found; and

o) second termination means coupled to said repetition means and said second comparison means for terminating said repetitions if no new lowest of the lowest accumulated scheduling cost is found for a predetermined number of consecutive repetitions.

14. The apparatus as set forth in claim 11, wherein said initial trial set of N weights comprises N randomly generated weights.

15. The apparatus as set forth in claim 12, wherein said second trial weight generation means systematically varies the immediately preceding trial set of N weights along an orthogonal dimension of a weight space formed by said weights, one orthogonal dimension at a time, by varying the weights of the immediately preceding trial set of N weights, one weight at a time, each weight being varied at least one time in a predetermined manner.

16. The apparatus as set forth in claim 15, wherein said second trial weight generation means varies each weight by adding a variation value to the weight being varied.

17. The apparatus as set forth in claim 16, wherein said second trial weight generation means computes each of said variation values by multiplying a current value of a first variable, a current value of a second variable and a current value of a third variable,

said second trial weight generation means varies said first variable from -1 to 1,

said second trial weight generation means varies said second variable incrementally in a predetermined manner from a first initial value over a first finite amount in increasing arithmetic increments, and

said second trial weight generation means varies said third variable incrementally in a predetermined manner from a second initial value over a second finite amount in fixed geometric increments.

18. The apparatus as set forth in claim 17, wherein said second generation means varies said first variable from 1 to 6 in increasing arithmetic increments of 2 and 3.

19. The apparatus as set forth in claim 17, wherein said second generation means varies said second variable from 1 to 1000 in fixed geometric increments of 10s.

20. The apparatus as set forth in claim 16 wherein said third generation means varies the immediately preceding trial set of N weights along the last orthogonal dimension with the last systematic variation made under said first manner by repeatedly adding the last variation value of said first manner to the last weight varied in said first manner of each immediately preceding trial set of N weights.