Документ взят из кэша поисковой машины. Адрес
оригинального документа
: http://www.adass.org/adass/proceedings/adass00/O3-03/
Дата изменения: Tue May 29 20:18:28 2001 Дата индексирования: Tue Oct 2 04:43:15 2012 Кодировка: Поисковые слова: vallis |
The Vector, Signal and Image Processing Library (VSIPL) is an open standard C language Application Programmer Interface (API) that allows portable and optimized single processor programs. OpenMP is an open standard C/Fortran API that allows portable thread based parallelism on shared memory computers. Both of these standards have enormous potential to allow users to realize the goal of portable applications that are both parallel and optimized.
Exploiting these new open standards requires integrating them into existing applications as well as using them in new efforts. Image processing is one of the key areas where VSIPL and OpenMP can have a large impact. Currently, a large fraction of image processing applications is written in the Interpreted Data Language (IDL) environment. The goal of this work is to show that it is possible to bring the performance benefits of these new standards to the image processing community in a high level manner that is transparent to users.
Wide area 2-D convolution is a staple of digital image processing (see Figure 1). The advent of large format CCDs makes it possible to literally ``pave'' with silicon the focal plane of an optical sensor. Processing of the large images obtained from these systems is complicated by the non-uniform Point Response Function (PRF) that is common in wide field of view instruments. This paper presents a fast, FFT based algorithm for convolving such images. This algorithm has been transparently implemented within IDL environment using VSIPL (for optimized single processor performance) with added OpenMP directives (for parallelism).
The inputs of image convolution with variable PRFs consist of a source image, a set of PRF images and a grid which locates the center of each PRF on the source image. The output image is the convolution of the input image with each PRF linearly weighted by its distance from its grid center. The computational basis of this convolution is addition with interpolation of 2-D overlapping FFTs (see Figure 2). Today, typical images sizes are in the millions (2K x 2K) to billions (40K x 40K) of pixels. A single PRF is typically thousands of pixels (100 x 100) pixels, but can be as small 10 x 10 or as large as the entire image. Over a single image a PRF will be sampled as few as once but as many as hundreds of times depending on the optical system.
There are many opportunities for parallelism in this algorithm. The simplest is to convolve each PRF separately on a different processor and then combine all the results on a single processor. This approach works well with VSIPL, OpenMP and IDL (see Figure 3). At the top level a user passes the inputs into an IDL routine that passes pointers to an external C function. Within the C function OpenMP forks off multiple threads. Each thread executes its convolution using VSIPL functions. The OpenMP threads are then rejoined and the results are added. Finally a pointer to the output image is returned to the IDL environment in the same manner as done by any other IDL routine.
This algorithm was implemented at Boston University, on an SGI Origin 2000 (64 300MHz MIPS 10000 processors with an aggregate memory of 16GB). IDL version 5.3 from Research Systems, Inc. was used along with SGI's native OpenMP compiler (version 7.3.1) and the TASP VSIPL implementation. Implementing the components of the system was the same as if each were done separately. Integrating the pieces (IDL/OpenMP/VSIPL) was done quickly, although care had to be taken to use the latest versions of the compilers and libraries. Once implemented the software can be quickly ported via Makefile modifications to any system that has IDL, OpenMP, and VSIPL (currently these are SGI, HP, Sun, IBM, and Red Hat Linux). We have conducted a variety of experiments that show linear speedups using different numbers of processors and different image sizes (see Figure 4). Thus, it possible to achieve good performance using open standards underneath existing high level languages.