LLVM Adding Support For IBM MASS Library For POWER Vectorization
A new addition to the LLVM code-base this week is initial support for IBM's MASS vectorization library.
IBM Mathematical Acceleration Subsystem (MASS) are a set of libraries with optimized versions of frequently-used mathematical functions and vectorized to make use of POWER hardware's high core/thread counts. IBM MASS is of similar nature to say Intel's Math Kernel Library (MKL). Those wanting to learn more about the MASS libraries can do so on IBM.com.
The LLVM code provides initial support for vectorization using the MASS vector library routines when using the -vector-library=MASSV for the LLVM compiler stack.
The review of this IBM MASS vectorization library optional support has been ongoing for a while but now merged. It will be interesting to benchmark when time allows on the Raptor Blackbird / Talos II to see if using this vectorization library makes a meaningful difference at this stage.
This addition is in the Git/SVN code for this fall's LLVM 9.0 release.
IBM Mathematical Acceleration Subsystem (MASS) are a set of libraries with optimized versions of frequently-used mathematical functions and vectorized to make use of POWER hardware's high core/thread counts. IBM MASS is of similar nature to say Intel's Math Kernel Library (MKL). Those wanting to learn more about the MASS libraries can do so on IBM.com.
The LLVM code provides initial support for vectorization using the MASS vector library routines when using the -vector-library=MASSV for the LLVM compiler stack.
The review of this IBM MASS vectorization library optional support has been ongoing for a while but now merged. It will be interesting to benchmark when time allows on the Raptor Blackbird / Talos II to see if using this vectorization library makes a meaningful difference at this stage.
This addition is in the Git/SVN code for this fall's LLVM 9.0 release.
1 Comment