Research on the Evaluation Criterion for Performance Searching Process of Self-Adapting Numerical Software

With the rapid development of computer hardware, researchers pay more and more attentions to how to make efficient use of computer resources and develop the high performance software. Traditionally, researchers produce manual-optimized software for a given machine or resort to the optimization of compiler. There are many problems in these two methods, including poor portability and the difficulty to synchronize with the development of hardware. When designing high performance self-adapting numerical software packages (SANS), such as FFTW, ATLAS, PHiPAC, OSKI, etc, using auto-tuning method becomes a very useful way and could avoid the problems mentioned above. However, the searching process always consumes too much time. In the experiment, it is found that the current best performance does not change rapidly during the whole process, which is completely the opposite of the change of performance gained at the runtime, and the best performance gained at the beginning is or almost close to the best performance gained after the whole search process. In order to search efficiently, a criterion Pt is proposed to evaluate the self-adapting searching process and the method on how to use the criterion to control the searching process for reducing the search time is also given. The experiment result shows that the method given can get an acceptable performance in a short time.