Stream Processors And FPGA's Heterogeneous Computing Research And Implementation

Academia and industry show a strong interest to stream processing technology for its high performance in compute-intensive applications. The reconfigurable technology gets people's recognition for its design flexibility and high performance per watt. These two technologies become two major research focuses in the area of computer architecture, and have a development in recent years.However, the applications in some special field also have a very rapid development which mainly reflected in two aspects. First, the scale of application is increasing. For example, in the field of video encoding and decoding, the resolution of video become more and more higher with the widespread of high-definition digital technology. Currently, the mainly used high-definition video resolution is 1920x1080, but the development of technology will inevitably lead the video's resolution to a new height. This means a rapid increase in computation of video encoding and decoding. Current processors are facing new challenges. Second, update of application's algorithm. The algorithm for each application is not changeless. With the continuous deepening of knowledge, the algorithm itself will have a new break through. If we customize a specific processor for an application, this processor will be the best in performance and power consumption. But once the new algorithm appears, this processor can not be used for it. Thus, the best situation is that the processor can dynamically change with the adjustment of application.Based on the two problems we have talked above, our study is to solve these two problems with the combination of stream processing technology and reconfigurable technology, so that we can get not only the performance of stream processor and also the flexibility of reconfigurable technology. Firstly, this issue describes the design methods and implementations of our heterogeneous computing platform from the perspective of computer architecture. After that, based on this platform we present a method of application development. Finally, combining with specific example, Motion JPEG, we further present how to develop an application and how to make the application run efficiently on our platform. We test our platform using M-JPEG with a 4096x2160 high-definition video sequence, and compare the test results with PC and Multi-DSP platform which also run this application. On our platform, encoding the 4096x2160 video sequences can reach 30.3 frame/s. This meets the requirements of real-time encoding. Compare with the other two platforms, our platform's performance is 10-50 times better than their. Fully reflects the computing power of our heterogeneous computing platform. Our platform has different application mapping methods according to different application requirements. This flexibility is also one of our design goals.