Gaussian convolutions computation is required in several scientific fields and, to this aim, efficient approximation methods, based on Recursive Filters (RFs), have been developed recently. Among them, Gaussian Recursive Filters (RFs) are designed to approximate the Gaussian convolution in a very efficient way. The accuracy of these methods, as is well known, can be improved by means of the use of the so-called K-iterated Gaussian recursive filters, that is in the repeated application of the basic RF. To improve the provided accuracy, K-iterated versions of these methods are also considered. Since it is often necessary to handle large size one-dimensional input signals, a parallel approach becomes mandatory. Recently, we proposed a parallel algorithm for the implementation of the K-iterated first-order Gaussian RF on multicore architectures. Here, using a similar parallelization strategy, based on a domain decomposition with overlapping, we propose a new implementation that would exploit, in terms of both accuracy and performance, the GPU (Graphics Processing Unit) capabilities on CUDA environment. Tests and experiments confirm the reliability and the efficiency of the proposed implementation.
A gaussian recursive filter parallel implementation with overlapping
Galletti A.;Marcellino L.
2019-01-01
Abstract
Gaussian convolutions computation is required in several scientific fields and, to this aim, efficient approximation methods, based on Recursive Filters (RFs), have been developed recently. Among them, Gaussian Recursive Filters (RFs) are designed to approximate the Gaussian convolution in a very efficient way. The accuracy of these methods, as is well known, can be improved by means of the use of the so-called K-iterated Gaussian recursive filters, that is in the repeated application of the basic RF. To improve the provided accuracy, K-iterated versions of these methods are also considered. Since it is often necessary to handle large size one-dimensional input signals, a parallel approach becomes mandatory. Recently, we proposed a parallel algorithm for the implementation of the K-iterated first-order Gaussian RF on multicore architectures. Here, using a similar parallelization strategy, based on a domain decomposition with overlapping, we propose a new implementation that would exploit, in terms of both accuracy and performance, the GPU (Graphics Processing Unit) capabilities on CUDA environment. Tests and experiments confirm the reliability and the efficiency of the proposed implementation.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.