Files and links (1)
PDFOpen Access (License Unspecified), Open Access
Thesis
A comparative performance analysis of the phase recovery algorithm for microstructure reconstruction
Master of Science (M.S.), Drexel University
Jun 2009
DOI:
https://doi.org/10.17918/etd-3082
Abstract
This thesis explores the high-performance implementation of a phase recovery algorithm for microstructure reconstruction of materials. Implementations on a variety of high-performance computing platforms, including multi-core and Graphics Processing Unit (GPU), were investigated and compared. The phase recovery algorithm is an iterative process requiring multiple Discrete Fourier Transform (DFT) computations each iteration. In order to achieve high-performance, it is necessary to use highly optimized fast Fourier transform (FFT) code to compute the DFTs. In our investigation, several FFT libraries, including FFTW, the Intel R Math Kernel Library (MKL), the CUFFT library for the NVIDIAR GPU, and the SPIRAL generated code, were used and compared. The SPIRAL system provides an extensible framework for generating and automatically optimizing implementations of DSP (digital signal processing) algorithms described using mathematical formulas, and is the most extensible of the platforms investigated here. The phas recovery algorithm intersperses FFT computations with point-wise computations, and while the FFTs are the dominant computation, the point-wise operations can have a significant impact on the overall performance. Therefore, simply relying on the performance of an optimized FFT library is insufficient to obtain optimal performance. Unlike the FFTW, MKL, and CUFFT libraries, the SPIRAL system allows the FFTs to be combined with the point-wise operations and the entire algorithm to be optimized. In this thesis, we obtained a mathematical formula representing the phase recovery algorithm that can be incorporated into the SPIRAL framework and utilize SPIRAL's parallel and vector code generation and optimization facilities. The SPIRAL code generated in this thesis is sequential. We estimate that with a vectorized and parallelized SPIRAL implementation, it is possible to obtain a 1.5-fold speedup for two-dimensional (2D) phase recovery and 1.88-fold speed up for 3D phase recovery over the MKL implementation.
Metrics
45 File views/ downloads
13 Record Views
Details
- Title
- A comparative performance analysis of the phase recovery algorithm for microstructure reconstruction
- Creators
- Anupama Shankar Kurpad - DU
- Contributors
- Jeremy Russell Johnson (Advisor) - Drexel University (1970-)Prawat Nagvajara (Advisor) - Drexel University (1970-)
- Awarding Institution
- Drexel University
- Degree Awarded
- Master of Science (M.S.)
- Publisher
- Drexel University; Philadelphia, Pennsylvania
- Resource Type
- Thesis
- Language
- English
- Academic Unit
- College of Engineering (1970-2026); Electrical (and Computer) Engineering [Historical]; Drexel University
- Other Identifier
- 3082; 991014632718604721