Performance Improvement of Gaussian Filter Using SIMD Technology

Maryam Moradifar, Asadollah Shahbahrami
The 11th Iranian and the first International Conference on Machine Vision and Image Processing (MVIP 2020)

Abstract

Denoising is an important process before applying other postprocessing techniques on medical images. To obtain better quality images many denoising approaches have been introduced. Gaussian filter is a spatial domain filter, which is proper to deblur and to remove noise from images. Since the Gaussian filter modifies the input signal by convolution with a Gaussian function it is a computationally intensive algorithm. Hence to enhance the performance of the algorithm, it is better to perform two 1-D convolution operations instead of one 2-D convolution operation and then parallelize it. In this paper in order to increase the performance of 1-D convolution operation, we exploit both Data- and Thread-Level Parallelism using parallel programming models such as Intrinsic Programming Model, Compiler’s Automatic Vectorization and Open Multi-Processing. The experimental results were shown that the performance of our implementations is much higher than other approaches. Performance improvements of Multi-threaded version of all implementations are significantly improved compared to single-core implementations, and a speedup of 52.33x obtained over the optimal scalar implementation.

Keywords: Gaussian Filter, 1-D Convolution, Single Instruction MultipleData, Data-level Parallelism, Thread-level Parallelism