Advancement of the Auto-SIMDization Technology in the XL Compiler

The Auto-SIMDization technology is constantly expanded to handle more challenging loops present in applications. As a common SIMD framework is leveraged across different platforms including BGL/P/Q, VSX/VMX, SPU, enhancements to one platform often directly benefit others. In this talk, a myriad of novel SIMD techniques implemented in the XL compiler in the past year will be presented. The techniques include SIMDization of non-stride-one memory access via SLP, identifying minimum and maximum locations in a SIMD loop, and SIMD for parallel loops. Loop kernels fr om real applications will be used to illustrate the techniques along with performance improvements. This talk will conclude with ideas for furthering the SIMD technology in the future.

Greg Steffan

Last modified: Tue Aug 26 09:20:53 EDT 2008