Evaluation of the Ti DaVinci DM3730 Coprocessors focusing on Image Processing Algorithms in order to disburden the CPU

Abstract

The conventional concept of single-processor hardware platforms is more and more replaced by systems using multiple processors like DSPs and GPUs, specialized for different tasks. This "diversification" among processing units is one of the reasons for the recent success of small and powerful mobile devices. Now, the challenge on multi-processor hardware architectures is to assign each task to the best fitting subsystem for a maximum of performance. But not only performance matters, usability and portability are as well important factors for the usage of coprocessors. This thesis evaluates the difference in using various processors, regarding common applications and especially image processing, based on the "Ti DaVinci DM3730 Digital Media Processor" architecture. At first, the available hardware is analysed, followed by benchmarks on concrete algorithms to gain insights about possible optimisation steps for each processor. Sample algorithms include vector addition, matrix multiplication, RGB to HSV colorspace conversion and creating downscaled images for Image-Pyramids. For optimisation, we take a look at cache access models like ”tiling”, image data arrangement in memory, "intrinsics", branch reduction and other typical bottlenecks. At the end, this leads to a recommendation how to assign tasks to each processor.