Applying Code Specialization to FFT Libraries for Integral Parameters

Code specialization is an approach that can be used to improve the sequence of optimizations to be performed by the compiler. The performance of code after specialization may vary, depending upon the structure of the application. For FFT libraries, the specialization of code with different parameters may cause an increase in code size, thereby impacting overall behavior of applications executing in environment with small instruction caches. In this article, we propose a new approach for specializing FFT code that can be effectively used to improve performance while limiting the code increase by incorporating dynamic specialization. Our approach makes use of a static compile time analysis and adapts a single version of code to multiple values through runtime specialization. This technique has been applied to different FFT libraries over Itanium IA-64 platform using icc compiler v 9.0. For highly efficient libraries, we are able to achieve speedup of more than 80% with small increase in code size.