High performance 2D convolution utilizing the AVX512 on a multi-core architecture

Abstract

Convolution is a time consuming operation, especially for signal and image processing, which led us to develop an efficient implementation of 2D convolution for a multi-core architecture utilizing AVX512 intrinsics and OpenMP. For single precision convolution, our algorithm is on average 2.30, 3.88, 5.75, and 19.95 times faster than the IPP, OpenCV, Baziotis's algorithm, and MKL libraries. For double precision convolution, our algorithm is on average 3.12, 5.10, and 16.95 times faster than the OpenCV, Baziotis's algorithm, and MKL libraries. We have also developed a hybrid 2D convolution algorithm, written in C and assembly, to further augment the processing speeds for small kernel sizes.

Keywords

• avx512
• advanced vector extension
• 2d convolution
• openmp
• x64 assembly language