Pattern Matching and Recognition 689 - Special Topics on Advanced Compiler Technologies

1
Pattern Matching and Recognition689 - Special
Topics on Advanced Compiler Technologies

• Alin Jula
• March,19th 2003

2
Motivation - Why do we care ?

• Abstract concepts are simpler to deal with and
  closer to the human language
• Enables implementation substitution
• e.g. recurrences can be replaced
• Eases maintenance process
• e.g. Polaris - 600,000 lines of code

3
General Issues

• Pattern matching and recognition Semantic
  compiler

Highest Level
Human languages English, Romanian, Spanish, etc
Higher - Level codes
STAPL
CASE Instruments
C, Java
Low - Level codes
Fortran,C
assembly
4
Example

• Given a set of uranium isotopes with random
  energies and velocities, calculate their
  coordinates (energy, velocities, space
  coordinates,etc) after 3 seconds of interaction
• Equivalent with thousands or maybe hundred of
  thousands of lines of code

5
Pattern Matching and Recognition
Program code
6
Pattern Matching and Recognition

• Program Understanding
• (Low to High Level)
• Code Generation and Replacement
• (High to Low Level)

7
Quilici - Overview

• Translates C into C
• Input C code in Abstract Syntax Tree form
• Output Programming plans ( then C )
• Analysis
• Bottom-Up on the Code
• Top-Down on the Programming Plans

8
Quilici - Example
Equivalent C code
C code
9
Quilici - Plan Library

• A Library of patterns (programming plans)
• Extends an existing Library (from Andersen
  Consulting for understanding Cobol).

10
Quilici - Programming Plan

• Programming Plan
• definition list of attributes
• recognition rule(s)
• components of the plan
• constraints on the components

11
Quilici -Programming Plan
Read Values within a while loop
12
Quilici - Library Organization

• Plan indices
• Connects plans hierarchically
• Objective narrows the search
• Specialization Constraints
• Stores only the ? constraints
• Objective eliminates commonalties
• Implied Plans
• Objects recognized form a particular plan
• Objective avoids matching them with the code,
  their existence is implied from the recognized
  plans

13
Quilici - The Algorithm

14
Quilici Example
15
Parallel Code
16
BH - Overview

• Parallelizes FORTRAN code using pattern matching
• Input (FORTRAN code - abstract syntax tree)
• Output ( parallel FORTRAN code )
• Analysis
• Patterns are found in the code
• User replaces these patterns with an already
  parallelized implementation

17
BH - Overview
18
BH- Pattern Language

• Describes
• programming concepts
• domain concepts
• The patterns are written in the pattern language
  before pattern recognition process

19
BH - Pattern Language Syntax
20
BH - Patterns Examples
21
BH - Pattern Classification

• Hierarchical approach
• Classification of patterns
• base-concept (atomic) - e.g. swap
• intermediate-concept - e.g. pivoting
• domain-concept - e.g. Gaussian elimination
• Dependencies between the patterns are stored in
  the library

22
BH - Pattern Classification
23
BH - Pattern Recognition

• The user is asked to specify the domain concepts
  as narrow as possible
• The AST is then compared against the Pattern
  Library

24
BH - Domain Patterns
25
BH - Algorithm

• Prune the space of patterns to search
• Base concepts are identified
• Intermediate patterns are built up from the base
  ones
• Domain concepts are built up from intermediate
  ones

• One Pass Search !
• Fast
• not Effective

26
BH - Algorithm
27
BH - Algorithm (cont.)
28
(No Transcript)
29
K - Overview

• Parallelizes FORTRAN 77 and C code using pattern
  matching
• Input (FORTRAN 77 and C code - abstract syntax
  tree)
• Output ( parallel C code )
• Analysis
• Graph based
• Hierarchical approach

30
K-Library

• Collected around of 150 patterns for scientific
  numeric codes
• Hierarchical approach
• e.g. matrix-multiplication loop over a dot
  product
• Patterns characterized based on the depth of the
  loop nest
• e.g. scalar 0. Dot product1,matrix
  multiplication2, etc

31
K - library
32
K- Library Organization

• Organized as a Pattern Hierarchy Graph(PHG)
• G(V,E) - V is a set of patterns, E set of
  dependencies between patterns
• m dependent of the trigger patterns mi, there is
  an edge between (m,mi)
• Pattern recognition becomes a path finding
  program in PHG

33
K - Library Organization
34
K - Algorithm
35
K - Paramat

• ParamatPARallelize Automatically by pattern
  MATching
• Tool for pattern recognition and algorithm
  substitution
• Automatic Parallelization Tool

36
K - Paramat
37
K -Paramat

• Algorithm replacement (recurrences)
• Data distribution
• Run time prediction

38
(No Transcript)
39
W Y - Overview

• Program Understanding as Constraint Satisfaction
  Problem
• a set of variables Xi
• a set of values for each Xi, Dom (Xi)
• set of constraints - permissible subsets to
  variable
• Constraints implemented in Prolog
• Searching algorithm - backtracking

40
W Y - Overview
41
W Y - Algorithm
42
W Y - Results
43
(No Transcript)
44
PAP -Overview

• Parallelizable Algorithmic Patterns
• Graphical tool - permits vizualization of the
  recognized concepts, together with the
  implementation within the program
• Integrated into Vienna Fortran Compilation System
  

45
PAP - Recognizer

• Performs a hierarchical parsing driven by concept
  recognition rules (implemented in Prolog)
• Concept
• its compositional hierarchy (set of composing
  subcomponents)
• Relationships and constraints among composing
  components

46
PAP - Recognizer

• PAP builds an Abstract Program Representation(APR)
  from a Basic Program Representation (BPR)
  (initial code )
• Base level of representation - Program Dependence
  Graph (PDG)
• nodes - statements
• edges
• control flow ( labeled with T/F)
• data dependence (labeled with variable
  identifier)
• The APR is a hierarchical PDG

47
PAP - APR
48
PAP - Traits

• PAP deals with
• Program variation - various implementations for
  the same algorithm
• Delocalization - implementation of a concept
  throughout the code
• Overlapping implementations - implementations of
  two or more concepts are merged

49
PAP - Output example
50
(No Transcript)
51
BFV - Overview

• Code Generator
• Graphical Tool
• Input
• Design Patterns
• Output
• C code.

52
BFV
53
BFV
54
BFV
55
Parallel Code
56
M Y - Algorithmic equivalence

• Let P and P be two subprograms. Define P and P
  to be algorithmically equivalent iff one can be
  obtained from the other one through following
  transformations
• Rename variables (bijective function-mapping)
• Reorder the statements without violating the
  semantics
• Permute the operands of any commutative operator

57
References

• Quilici "An Opportunistic, Memory-Based
  Approach to Recognizing Programmings Plans" ,
  Alex Quilici
• BH "A Pattern Matching Approach for reusing
  Software Libraries in Parallel Systems" ,S.
  Bhansali, J.R. Hagemeister
• K "Pattern-Driven Automatic Parallelization" ,
  Christoph Kessler
• WY "Program Understanding as Constraint
  Satisfaction Representation and Reasoning
  Techniques" , S. Woods, Q. Yang
• PAP"PAP Recognizer a Tool for Automatic
  Recognition of Parallelizable Patterns" , B. Di
  Martino, G. Iannello
• BFV "Automatic Code Generation from Design
  Patterns, IBM Systems Journal, Vol. 35, No. 2,
  Frank Budinsky, Marilyn Finnie, Patsy Yu,John
  Vlissides
• MY "Automatic Algorithm Recognition and
  Replacement , Robert Metzger and Zhaofang Wen