Two Dimensional (2D) Image Convolution in CUDA by Shared & Constant Memory: An Optimized way

          After learning the concept of two dimension (2D) Convolution and its implementation in C language; the next step is to learn to optimize it. As Convolution is one of the most Compute Intensive task in Image Processing, it is always better to save time required for it. So, today I am going to share a technique to optimize the Convolution process by using CUDA. Here we will use Shared Memory and Constant Memory resources available in CUDA to get fastest implementation of Convolution.

Two Dimensional (2D) Image Convolution : A Basic Approch

Image Convolution is a very basic operation in the field of Image Processing. It is required in many algorithms in Image Processing. Also it is very compute intensive task as it involves operation with pixels.

One Dimensional (1D) Image Convolution in CUDA by using TILES

          Tiled algorithms are a special case in CUDA as we can Optimize the algorithm implementation, by using this strategy. It is very useful when we want to achieve maximum usage of available GPU hardware, present in the system. It has several advantages over naive CUDA implementations such as improved Memory bandwidth, reduced memory read/write operations,etc. Tiled implementation uses Shared memory available in GPU hardware which is much faster as compared to Global Memory in GPU. In any naive CUDA implementation only Global memory is used for all read and write operations. So, if these memory (read/write) operations are huge in number then the more time is wasted only in transferring the data which results in low/poor performance.

One Dimensional (1D) Image Convolution in CUDA

          First let me tell you that if you are reading this page then you are already looking for some advance stuff in today's technology as both CUDA & Image Processing are highly demanding as well as advanced technologies. On this blog we will be mainly focusing on use of CUDA(Compute Unified Device Architecture) technology to improve Image Processing Algorithms.The improvement is mainly with respective to Time & Space required to Image Processing Algorithm. You may refer to concern links provided to get more information about both the fields.