Method of, system for, and computer program product for performing weighted loop fusion by an optimizing compiler

• US 6,058,266 A
• Filed: 06/24/1997
• Issued: 05/02/2000
• Est. Priority Date: 06/24/1997
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Megiddo, Nimrod, Sarkar, Vivek
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Chaki, Kakali
• Application Number:
  US08/881,883
• Time in Patent Office: 1,043 Days
• Field of Search: 395/705, 395/706, 395/708, 395/709, 395/588
• US Class Current: 717/156
• CPC Class Codes: G06F 8/443 Optimisation
• CPC Class Codes: G06F 8/452 Loops

Abstract

An integer programming formulation for weighted loop fusion is presented. Loop fusion is a well-known program transformation that has shown to be effective in reducing loop overhead and improving register and cache locality. Weighted loop fusion is the problem of finding a legal partition of loop nests into fusible clusters so as to minimize the total inter-cluster edge weights. Past work has shown that the weighted loop fusion problem is NP-hard. Despite the NP-hardness property, the present invention provides optimal solutions that may be found efficiently, in the context of an optimizing compiler, for weighted loop fusion problem sizes that occur in practice. An integer programming formulation for weighted loop fusion with a problem size (number of variables and constraints) that is linearly proportional to the size of the input weighted loop fusion problem is also presented. The integer programming formulation may be solved efficiently using a general integer programming package. Alternatively, a custom branch-and-bound procedure for the integer programming formulation is presented that is more efficient than the procedures used in general integer programming.

Claims

1. A method of performing weighted loop fusion on a plurality of loop nests by a computer system, said method comprising the steps of:

creating a loop dependence graph (LDG) representing the plurality of loop nests, wherein each of the plurality of loop nests is associated with a node of the LDG, and wherein an edge of the LDG represents a loop-independent data dependence from a source loop nest to a destination loop nest;

associating a weight w_ij =w_ji with each pair of nodes i and j, wherein the weight represents a cost savings obtained if loop nests i and j are fused;

determining a set, AN×

N, of directed summary arcs on N, wherein N is a set of nodes of the loop dependence graph, N={1, 2, . . . , n};

determining a set, C, of contractable arcs wherein C is a subset of the set of directed summary arcs, CA;

determining a set, NC, of noncontractable arcs wherein NC is a subset of the set of directed summary arcs, NC=A-C

determining a set, B, of unordered pairs of nodes (i,j) that have a w_ij >

0 and that are not connected by a summary arc, both (i,j).epsilon slash.A and (j,i).epsilon slash.A;

associating a (0,1)-variable x_ij with each unordered pair of nodes (i,j), and setting the x_ij =0 if and only if i and j are placed in a same cluster and setting x_ij =1 otherwise;

determining a set of multipliers π

_i (i.di-elect cons.N) wherein each of the multipliers π

_i are an integer between 0 and n=|N|, wherein π

_j -π

_i ≧

1 for all (i,j).di-elect cons.A, wherein π

_i =π

_j if x_ij =0, wherein π

_j -π

_i ≧

1 if x_ij =1 and (i,j).di-elect cons.A; and

minimizing a summation of a product of w_ij and x_ij over a set of nodes (i,j) comprising a union of the set B and the set of contractable arcs C, (i,j).di-elect cons.B∪

C, subject to a set of constraints comprising;

##EQU10## performing loop fusion on the plurality of loop nests using loop partitions based on the minimized summation.

2. The method of claim 1 wherein the minimizing step comprises minimizing a summation of a product of w_ij and x_ij over a set of nodes (i,j) comprising a union of the set B and the set of contractable arcs C, (ij).di-elect cons.B∪

C, subject to a set of constraints comprising;

##EQU11##

3. The method of claim 2 further comprising the steps of:

arranging subproblems in the form of a branch and bound tree wherein a node (C'"'"',O) of the branch and bound tree is characterized by two disjoint subsets;

(i) C'"'"'C, a set of arcs have been decided to remain not contracted, and

(ii) OC, a set of arcs that have already been contracted;

defining b_ij =0 if and only π

_i =π

_j, and b_ij =1 otherwise;

for each subproblem, minimizing a summation of a product of w_ij and b_ij over a set of nodes (i,j).di-elect cons.B∪

C-(C'"'"'∪

O), subject to a set of constraints comprising;

##EQU12## determining a lower bound for a node (C'"'"',O) of the branch and bound tree by determining V₁ (C'"'"',O)≧

[V₂ (C'"'"',O)]/n'"'"'+Σ

.sub.(i,j).di-elect cons.O w_ij, wherein an optimal value of (C'"'"',O)₁ is V₁ (C'"'"',O), an optimal value of (C'"'"',O)₂ is V₂ (C'"'"',O), and n'"'"' is a number of nodes remaining after a contraction given by O; and

deriving an upper bound for a node (C'"'"',O) of the branch and bound tree by determining V₁ (C'"'"',O)≦

V₂ (C'"'"',O)+Σ

.sub.(i,j).di-elect cons.O w_ij.

4. An article of manufacture for use in a computer system for performing weighted loop fusion on a plurality of loop nests by the computer system, said article of manufacture comprising a computer-readable storage medium having a computer program embodied in said medium which may cause the computer system to:

create a loop dependence graph (LDG) representing the plurality of loop nests, wherein each of the plurality of loop nests is associated with a node of the LDG, and wherein an edge of the LDG represents a loop-independent data dependence from a source loop nest to a destination loop nest;

associate a weight w_ij =w_ji with each pair of nodes i and j, wherein the weight represents a cost savings obtained if loop nests i and j are fused;

determine a set, AN×

N, of directed summary arcs on N, wherein N is a set of nodes of the loop dependence graph, N={1, 2, . . . , n};

determine a set, C, of contractable arcs wherein C is a subset of the set of directed summary arcs, CA;

determine a set, NC, of noncontractable arcs wherein NC is a subset of the set of directed summary arcs, NC=A-C

determine a set, B, of unordered pairs of nodes (i,j) that have a w_ij >

0 and that are not connected by a summary arc, both (i,j).epsilon slash.A and (j,i).epsilon slash.A;

associate a (0,1)-variable x_ij with each unordered pair of nodes (i,j), and setting the x_ij =0 if and only if i and j are placed in a same cluster and setting x_ij =1 otherwise;

determine a set of multipliers π

_i (i.di-elect cons.N) wherein each of the multipliers π

_i are an integer between 0 and n=|N|, wherein π

_j -π

_i ≧

1 for all (i,j).di-elect cons.A, wherein π

_i =π

_j if x_ij =0, wherein π

_j -π

_i ≧

1 if x_ij =1 and (i,j).di-elect cons.A; and

minimize a summation of a product of w_ij and x_ij over a set of nodes (i,j) comprising a union of the set B and the set of contractable arcs C, (i,j).di-elect cons.B∪

C, subject to a set of constraints comprising;

##EQU13## perform loop fusion on the plurality of loop nests using loop partitions based on the minimized summation.

5. The article of manufacture of claim 4 wherein the computer program in causing the computer system to minimize the summation may further cause the computer system to minimize a summation of a product of w_ij and x_ij over a set of nodes (i,j) comprising a union of the set B and the set of contractable arcs C, (i,j).di-elect cons.B∪

C, subject to a set of constraints comprising;

##EQU14##

6. The article of manufacture of claim 5 wherein the computer program may further cause the computer system to:

arrange subproblems in the form of a branch and bound tree wherein a node (C'"'"',O) of the branch and bound tree is characterized by two disjoint subsets;

(i) C'"'"'C, a set of arcs have been decided to remain not contracted, and

(ii) O∪

C, a set of arcs that have already been contracted;

define b_ij =0 if and only π

_i =π

_j, and b_ij =1 otherwise;

for each subproblem, minimize a summation of a product of w_ij and b_ij over a set of nodes (i,j).di-elect cons.B∪

C-(C'"'"'∪

O), subject to a set of constraints comprising;

##EQU15## determine a lower bound for a node (C'"'"',O) of the branch and bound tree by determining V₁ (C'"'"',O)≧

[V₂ (C'"'"',O)]/n'"'"'+Σ

.sub.(i,j).di-elect cons.O w_ij, wherein an optimal value of (C'"'"',O)₁ is V₁ (C'"'"',O), an optimal value of (C'"'"',O)₂ is V₂ (C'"'"',O), and n'"'"' is a number of nodes remaining after a contraction given by O; and

derive an upper bound for a node (C'"'"',O) of the branch and bound tree by determining V₁ (C'"'"',O)≦

V₂ (C'"'"',O)+Σ

.sub.(i,j).di-elect cons.O w_ij.

7. A computer system for performing weighted loop fusion on a plurality of loop nests by the computer system, said computer system comprising:

means for creating a loop dependence graph (LDG) representing the plurality of loop nests, wherein each of the plurality of loop nests is associated with a node of the LDG, and wherein an edge of the LDG represents a loop-independent data dependence from a source loop nest to a destination loop nest;

means for associating a weight w_ij =w_ji with each pair of nodes i and j, wherein the weight represents a cost savings obtained if loop nests i and j are fused;

means for determining a set, AN×

N, of directed summary arcs on N, wherein N is a set of nodes of the loop dependence graph, N={1, 2, . . . , n};

means for determining a set, C, of contractable arcs wherein C is a subset of the set of directed summary arcs, CA;

means for determining a set, NC, of noncontractable arcs wherein NC is a subset of the set of directed summary arcs, NC=A-C

means for determining a set, B, of unordered pairs of nodes (i,j) that have a w_ij >

0 and that are not connected by a summary arc, both (i,j).epsilon slash.A and (j,i).epsilon slash.A;

means for associating a (0,1)-variable x_ij with each unordered pair of nodes (i,j), where the x_ij =0 if and only if i and j are placed in a same cluster and x_ij =1 otherwise;

means for determining a set of multipliers π

_i (i.di-elect cons.N) wherein each of the multipliers π

_i are an integer between 0 and n=|N|, wherein π

_j -π

_i ≧

1 for all (i,j).di-elect cons.A, wherein π

_i =π

_j if x_ij =0, wherein π

_j -π

_i ≧

1 if x_ij =1 and (i,j).di-elect cons.A; and

means for minimizing a summation of a product of w_ij and x_ij over a set of nodes (i,j) comprising a union of the set B and the set of contractable arcs ##EQU16## means for performing loop fusion on the plurality of loop nests using loop partitions based on the minimized summation.

8. The computer system of claim 7 wherein the minimization further comprises a minimization of a summation of a product of w_ij and x_ij over a set of nodes (i,j) comprising a union of the set B and the set of contractable arcs C, (i,j).di-elect cons.B∪

C, subject to a set of constraints comprising;

##EQU17##

9. The computer system of claim 8 further comprising:

an arrangment of subproblems in the form of a branch and bound tree wherein a node (C'"'"',O) of the branch and bound tree is characterized by two disjoint subsets;

(i) C'"'"'C, a set of arcs have been decided to remain not contracted, and

(ii) OC, a set of arcs that have already been contracted;

a set b_ij where b_ij =0 if and only π

_i =π

_j, and b_ij =1 otherwise;

a minimization for each subproblem of a summation of a product of w_ij and b_ij over a set of nodes (i,j).di-elect cons.B∪

C-(C'"'"'∪

O), subject to a set of constraints comprising;

##EQU18## a determination of a lower bound for a node (C'"'"',O) of the branch and bound tree by determining V₁ (C'"'"',O)≧

[V₂ (C'"'"',O)]/n'"'"'+Σ

.sub.(i,j).di-elect cons.O w_ij, wherein an optimal value of (C'"'"',O)₁ is V₁ (C'"'"',O), an optimal value of (C'"'"',O)₂ is V₂ (C'"'"',O), and n'"'"' is a number of nodes remaining after a contraction given by O; and

a determination of an upper bound for a node (C'"'"',O) of the branch and bound tree by determining V₁ (C'"'"',O)≦

V₂ (C'"'"',O)+Σ

.sub.(i,j).di-elect cons.O w_ij.