Telescoping Languages

1
Telescoping Languages

• This presentation will probably involve audience
  discussion, which will create action items. Use
  PowerPoint to keep track of these action items
  during your presentation
• In Slide Show, click on the right mouse button
• Select Meeting Minder
• Select the Action Items tab
• Type in action items as they come up
• Click OK to dismiss this box
• This will automatically create an Action Item
  slide at the end of your presentation with your
  points entered.

• Bradley Broom, Ken Kennedy et al.
• Rice University

2
The problem?

• Getting high performance on modern computer
  architectures requires
• lots of programming
• machine specific tuning
• Not enough programmers with the necessary skills.
• Would still like faster development.

3
The Solution

• Lines of code per day per programmer essentially
  constant.
• Use more productive languages.
• Domain specific languages
• COBOL vs. SQL
• User interface generators
• Network protocol stub generators
• lex/yacc scanner parser generators

4
Languages for niche domains

• Small community of interest cant justify major
  language development.
• A domain-specific library in effect defines a new
  language.
• Two ways to support new syntax
• interpretation
• compilation via simple rewriting.

5
Problem with libraries

• Very high-level routines can be inefficient
• For instance calling two routines that each
  touch all elements of an array.
• Poor memory hierarchy utilization.
• A number of other implementations may only need a
  single pass over the array.
• Providing a large number of variants for each
  routine is not the answer.

6
Compiler Optimization

• Whole program optimization might possibly
  generate efficient code.
• But much too slow
• Depends on size of whole program, including all
  the libraries.
• Still doesnt incorporate expert knowledge about
  the problem domain.
• Use of specialized algorithms if special
  properties apply (e.g. two matrices are both
  upper-triangular).

7
Telescoping Languages

• Give compiler ability to optimize library calls
  just like primitive operations.
• Properties of library calls,
• Relationships between variants of library
  operations,
• Program transformations likely to improve
  performance.
• Do this in a generic way, just once.

8
TeleGen

• Telescoping Optimizer Generator
• Inputs
• Library source
• Annotations
• (by library developer or other expert)
• Outputs
• Augmented library source
• compiled by vendor compiler
• contains generated library variants
• Library aware optimizer

9
Manual Studies

• KeLPIO
• Signal-processing in Matlab
• Matlab library optimizer
• Signal processing library optimizer.
• Generation of Fortran library from Matlab master
  source.

10
Hand Studies continued

• Biostatistical modeling in S-Plus
• Code used in modeling cancer trials
• Developed in S-Plus
• Rewritten in C/Fortran to evaluate properties via
  simulation.
• A factor of 250 improvement by generic S-Plus
  library optimization.
• A further factor of 90 if domain specific array
  properties are known.

11
Under development at Rice

• Matlab / KeLPIO specific optimizers.
• Generic Fortran infrastructure
• Parser based on SGI Open64 compiler.
• Analysis based components in Rice dHPF compiler.
• New annotation language.
• New telescoping language generator.

12
Could you use Telescoping Languages?

• Bradley Broom ltbroom_at_rice.edugt