Study on Auto-tuning Method for Numerical Computation and Embedded Systems

在本研究中，我们将传统的数值计算库自动调优方法扩展到更广泛的数值软件自动调优方法，并展示了嵌入式系统开发过程中自动调优的可能性。本文的研究结果可以概括为以下四点：(1)自调优算法的研究结果：(a)对于数值库来说，如何为问题大小N设置自调优系统上的采样点是至关重要的。在本研究中，我们开发了一种新的自动调谐方法，使自动调谐系统可以在运行时改变采样点。在密集矩阵库和稀疏矩阵库上评价了该方法的有效性。(b) UC最近提出了一种计算对称密集特征值和特征向量的新方法。伯克利。我们提出了一种基于MRRR的库上双节例程的运行时自动调优方法。(2)自调优语言的研究结果：对于提出的ABCLibScript自调优语言，我们讨论了将ABCLibScript函数扩展到嵌入式系统的可能性。当我们为当前的ABCLibScript技术应用自动调优方法时，我们展示了扩展的解决方案和实现线索。(3)自动调优方法在嵌入式系统上的扩展：我们验证了嵌入式系统自动调优的初步性能。该应用程序使用了一个数值计算基准，如信号处理，用于嵌入式系统。因此，我们使用基准测试获得了6.4倍的加速。(4)建立自调优与嵌入式系统的小型研讨会：我们邀请了一位研发部门的研究员到某公司进行了一个现场小型研讨会，展示和讨论嵌入式系统开发过程中遇到的实际问题。本次研讨会的影响必然会影响到自动调优研究的新要求。

In this research, we extend conventional auto-tuning method for numerical computation libraries to more widely-covered method for numerical software and to show possibility of auto-tuning for the development process of embedded systems.The results of this research are summarized as the following four points :(1) Results on Auto-tuning Algorithm : (a) It is crucial that how to set the sampling points on auto-tuning system for N, which is problem size, for numerical libraries. In this research, we develop new auto-tuning method that auto-tuning system can change the sampling points at run-time. The effectiveness of the proposed method was evaluated on dense and sparse matrix libraries. (b) A new method for computing symmetric dense eigenvalues and eigenvector computation are recently proposed at UC. Berkeley. The method is named MRRR We proposed an auto-tuning method at run-time for the routine of bi-section on the library using MRRR.(2) Results on Auto-tuning Language: For the proposed auto-tuning language of ABCLibScript, we discussed the possibility of extension for ABCLibScript functions to embedded systems. We showed the solution and implementation cues for the extension when we apply the auto-tuning method for current ABCLibScript technologies.(3) Towards to Extension of Auto-Tuning Method on Embedded Systems : We verified the preliminary performance on auto-tuning for embedded systems. The application was used a numerical computation benchmark, like signal processing, for embedded systems. As a result, we obtained 6.4x speedups using a benchmark.(4) Establishment of a Mini Workshop for Auto-tuning and Embedded Systems : We took a place mini workshop by inviting a researcher for research and development division on a company, to show and discuss the real problems on the development process of embedded systems. The impact of this workshop must affect new requirements of auto-tuning study.