The platform-dependent database plays a key role in high-level performance estimation, however issues regard to the database construction are rarely studied. Therefore in this paper, a slope-threshold method is proposed to determine the number of sufficient samples to construct a reliable and accurate platform-dependent database. Besides, to improve estimation accuracy further, we also proposed a node refinement method which decomposes coarsely partitioned subtask into smaller ones. To verify the proposed approaches, MPEG-4 shape coding is taken as an example. The result shows that using the database constructed with the slope-threshold method improves both estimation accuracy and error estimation coverage, especially when subtasks are coarsely partitioned. If node refinement is also incorporated, the estimation error percentage can be improved from 3.20% to 1.85%.
|Number of pages||4|
|State||Published - 1 Dec 2004|
|Event||2004 IEEE Asia-Pacific Conference on Circuits and Systems, APCCAS 2004: SoC Design for Ubiquitous Information Technology - Tainan, Taiwan|
Duration: 6 Dec 2004 → 9 Dec 2004
|Conference||2004 IEEE Asia-Pacific Conference on Circuits and Systems, APCCAS 2004: SoC Design for Ubiquitous Information Technology|
|Period||6/12/04 → 9/12/04|