Memory Efficient Deployment Of An Optical Flow Algorithm On Gpu Using Openmp

In this paper, we consider the recent set of OpenMP directives related to GPU deployment and seek an evaluation through the case of an optical flow algorithm. We start by investigating various agnostic transformations that attempt to improve memory efficiency. Our case study is the so-called Lucas-Kanade algorithm, which is typically composed of a series of convolution masks (approximation of the derivatives) followed by 2 x 2 linear systems for the optical flow vectors. Since, we are dealing with a stencil computation for each stage of the algorithm, the overhead of memory accesses together with the impact on parallel scalability are expected to be noticeable, especially with the complexity of the GPU memory system. We compare our OpenMP implementation with an OpenACC one from our previous work, both on a Quadro P5000.