This is a helper library for writing fast x86 SIMD kernels.
This library is currently in an "initial hacking" stage, and will be hard to use until it gets more documentation! Let me know if this should be a higher priority.
Authors:
Compilation instructions:
# First you need to create a file Makefile.local, which defines Makefile variables
# INCDIR install dir for C++ headers
# CPP c++ compiler command line (should probably include flags -std=c++11 -O3 -march=native)
#
# You can probably just symlink Makefile.local to one of the following:
# site/Makefile.local.norootprivs (specifies INCDIR=$HOME/include)
# site/Makefile.local.rootprivs (specifies INCDIR=/usr/local/include)
# Assuming you're not installing as root
ln -s site/Makefile.local.norootprivs Makefile.local
make test # optional but recommended
make install # Installs .hpp files to $(INCDIR)
Note 1: To use the simd_helpers library, you always need to compile with -march=native,
otherwise compilation will fail with a long list of errors.
If you're using gcc, I find that the flag --param inline-unit-growth=10000 sometimes helps.
This allows more aggressive inlining. (Example: the CHIME FRB search pipeline runs significantly
faster with this flag.)
Note 2: You may find that compilation fails, with the error message below, followed by a very long list of compiler errors.
Either you're compiling on an old machine, or you forgot the -march=native compiler flag.
If this is an old machine, see note in simd_helpers/README.md.
This error message is generated if your CPU appears to not have the SSE 4.2 instruction set.
If you're using a machine more recent than 2008, then you probably just need to include -march=native
in your compiler flags.
If you are getting this error because you have an old (pre-2008) machine, then send me an email and it will probably be easy for me to get simd_helpers working on your machine. The only reason I haven't done it yet is that I don't have access to a machine old enough to use for testing!