Profile-directed Dynamically Loading of Program into Scratch-pad Memory

• Main Authors: I-Jui Sung, 宋宜叡
• Other Authors: Chung-Ping Chung
• Format: Others
• Language: en_US
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/64563300131731962703

碩士 === 國立交通大學 === 資訊工程系所 === 92 === Effective utilization of on-chip memories has always been an important factor to improve the performance of a program. Caching techniques has been studied for decades, whilst recently scratch-pad memory is getting more and more focuses on it, because of its lower cost and lower power-consumption comparing to the cache. Due to the nature of scratch-pad memory, it is required to modify current program to effeciently utilize the scratch-pad memory. Hence, for legacy programs, compiler researchers have proposed automatic modification methods of source program in compilers supporting scratch-pad memory. In this thesis we focus on the dynamic loading of parts of program to on-chip scratchpad memory to improve the performance. Previous dynamic approach applies the ILP framework to solve the problem deciding optimal set of dynamically-loaded program parts, but not feasible when program size grows. Hence, a heuristic to solve the problem of deciding dynamically loading parts of a program into scratch-pad memory is proposed and evaluated in this thesis. Our heuristic is based on the idea of ‘divide-and-conquer’, here ‘divide’ corresponds to the decomposition of nested loops and ‘conquer’ corresponds to an one-dimensional 0/1 knapsack-packing method. Evaluation results showed that our heuristic outperformed the static approach, and the performance of software-controlled dynamically-loaded code in scratch-pad memory produced by our method is comparable to traditional I-cache. This suggests a viable path for larger applications to efficiently utilize the scratch-pad memory.