Automated Specific Instruction Customization Methodology for Multimedia Processor Acceleration

Kang Zhao1,  Jinian Bian1,  Sheqin Dong1,  Yang Song2,  Satoshi Goto2
1Tsinghua University, China, 2Waseda University, Japan


Abstract

To enhance the computing ability of the multimedia processor, this paper presents an automated specific instruction customization methodology which integrates multiple processes such as profiling, specific instruction generation, and processor synthesis. Especially, this methodology features a profiling stage which is equipped with a sub-graph matching algorithm using isomorphism analysis. Furthermore, to support the features of the multimedia applications, three special structures are integrated in the instruction customization, such as fusion instructions, parallel and pipelining structures. Finally the new instructions are inserted into the processor core and then synthesized via an instruction language called TIE. To evaluate this methodology, a case study on the intra prediction of H.264 encoder is adopted, and a H.264-JM8.0 system is built based on the Xtensa toolset from Tensilica Inc. We have verified this system with a set of commonly used benchmarks, and compared the results with the system without specific instructions. The final experimental results indicate that 67% enhancement can be obtained in accelerating the computation via specific instructions.