ARM NEON improvements in Eigen 3.3

Question

I am using Eigen C++ library to write certain algorithm for 64-bit ARM Cortex-A53 processor.

It says in the Eigen changelog (http://eigen.tuxfamily.org/index.php?title=3.3) that there are ARM NEON improvements in release 3.3, including "correct tuning of the target number of vector registers." Now, does anybody know what does this actually mean? Is it possible to tell Eigen to take full advantage of the 32 vector registers of A53? According to my quick tests it didn't happen automatically.

Thanks, Lauri

Answer 1

Thanks for the report... technically, you found a bug!

Eigen should tune its usage of NEON 128-bit registers to that in arm64 mode, it uses 32 of them. I thought that I had made it do that, but it looks like I was wrong :-/ https://bitbucket.org/eigen/eigen/src/tip/Eigen/src/Core/arch/NEON/PacketMath.h?fileviewer=file-view-default#PacketMath.h-34

One should fix that...