|
|
| |
| Creating a Compiler Optimized Inlineable Implementation of Intel Svml Simd Intrinsics |
Promyk Zamojski, Raafat Elfouly
|
Abstract:
Abstract: Single Input Multiple Data (SIMD) provides data parallelism execution via implemented SIMD instructions and registers. Most mainstream computers architectures have been enhanced to provide data parallelism though SIMD extensions. These advances in parallel hardware have not been accompanied by the necessary software libraries granting programmers a set of functions that could work across all major compilers adding a 4x, 8x or 16x performance increase compared to standard SISD. Intel’s SVML library offers SIMD implementation of Math and Scientific functions that have never been ported to work outside of Intel's own compiler. An Open Source inlineable implementation would increase performance and portability. This paper illustrates the development of an alternative compiler neutral implementation of Intel's SVML library.
|
|
|
|