A. Kibkalo, M. Lotkov, I. Rogozhkin, and A. Turovets (Russia)
Image manipulation, performance library, data pipeline
A new approach to fast processing of big images is proposed. The image size now often exceeds the size of CPU L2 cache. While performing operations on such images data access takes generally more time than operation itself. There are two main methods to accelerate calculations in this case. First – development of optimized image processing libraries containing fast implementations of main image processing algorithms (e.g. Intel IPP). Second – description of the task in a some abstract form and organization of delayed pipelined processing of the image by small fragments. We combined these two approaches in the new Deferred Mode Image Processing library (DMIP). It is the software layer over Intel IPP providing formula-level programming, pipelined processing of images and easy extension with user-defined operations. DMIP library allows for accelerating of image processing tasks up to 3 times.
Important Links:
Go Back
