ABCLib Project:
The project of development for ABCLib : A Parallel Numerical Library with
Auto-tuning Facility
Online Manual
Updated: October 2005
Contents
BLAS | |
An Auto-tuning Facility for BLAS3: AutoTuned-RB |
|
AutoTuned-RB |
Linear Equations | |
A Run-time Auto-tuning Facility for Sparse Solvers: RAO-SS |
|
RAO-SS |
Eigensolver | |
A Eigensolver for Dense Real Matrices: ABCLib_DRSSED |
|
ABCLib_DRSSED |
Benchmark | |
A Benchmark Using Numerical Library with Auto-tuning Facility: ABCLibBench |
|
ABCLibBench |
Language To Support Auto-tuning Facility | |
A Langage To Support Development of Numerical Library with Auto-tuning
Facility: ABCLibScript |
|
ABCLibScript |
Visualizer | |
A Visualizer To Support Debelopment of Numerical Library with Auto-tining
Facility: VizABCLin |
|
VizABCLib (Opened in October 7,2005) |
Foreword
This manual describes the usage of the Dense Real Symmetric Standard Eigenvalue Decomposition (ABCLib_DRSSED) developed as part of the Automatically Blocking and Communication-adjustment Library (ABCLib) project .
ABCLib is a linear mathematics library that achieves high performance in parallel-processor computers having hierarchical memory as processing elements (PEs). Specifically, ABCLib is a parallel-computing library designed to deliver high performance in PC clusters, in which inexpensive PCs are networked as PEs, as well as on supercomputers.
The ABCLib project is developing this library under the following guiding principles:
Active use of blocking algorithms in order to achieve high-performance numerical calculation in computers having hierarchical memory.
Active use of communications encapsulation algorithms, in order to achieve high performance in slow networking environments.
Inclusion of auto-tuning features to support a wide range of parallel-processing environments, from PC clusters to supercomputers.
The inclusion of auto-tuning features makes the library easy to use, by vastly reducing the number of user-specified parameters over previous libraries. In addition, the auto-tuning features are implemented in such a way that having fewer parameters does not degrade performance.
Directives are provided to support additional auto tuning, for developers of mathematics libraries with auto-tuning features.
Benchmarking software using the auto-tuning framework is provided in order to evaluate the performance of the compiler, system, etc.
The following linear-calculation functions are planned to be added to the ABCLib.
For dense matrices:
Direct solution of simultaneous linear equations
Direct solution of eigenvalue problems
Iterative solution of eigenvalue problems
For sparse matrices:
Direct solution of simultaneous linear equations
Iterative solution of simultaneous linear equations
Iterative solution of eigenvalue problems
The following features are also provided.
A tool to support the addition of auto-tuning features:
ABCLibCodeGen: A preprocessor for ABCLibScript directives to support the addition of
auto-tuning features using the FIBER method
Benchmarking software:
ABCLibBench: Benchmarking using ABCLib
The purpose of the release of this test version is to enable its use by many different users, in order to improve its utility, find and correct bugs, and provide feedback for new development. Consequently, please be aware that there may still be problems with the library, including deficiencies in the manual, poor code readability, and sudden changes in specifications.
Please feel free to contact the author regarding any questions or comments you may have.