Fast Multiscale Algorithms for Information Representation and Fusion
Abstract
In the second quarter of the work effort, we continued research and development of algorithms based on randomized matrix decompositions. These randomized variants have theoretically proven improvements in computational complexity over existing algorithms. Algorithm designs for computing the Randomized Singular Value Decomposition (SVD) using randomized Fast Fourier Transform projections and the Interpolative Decomposition were completed. Fortran 95 interfaces for reusable randomized SVD routines have been defined and implemented. A survey of currently available optimized libraries for the Basic Linear Algebra Subprograms (BLAS) and Linear Algebra PACKage (LAPACK) interfaces was conducted to guide development. The randomized SVD implementation uses these libraries via standardized interfaces to make optimal use of hardware resources (e.g., multiple cores, CPU cache) in addition to using the OpenMP standard (for parallel execution of code). Use of these standards enables the code to be built flexibly in a number of ways on various target platforms. Preliminary testing of the software is complete. Additional updating, fine tuning will be based on results from various experiments that will be conducted in the upcoming quarters.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 2011
- Accession Number
- ADA538312
Entities
People
- Devasis Bassu