Any insight on how AVX-512 will effect Julia?

[Eric Diaz]

Just found out about AVX-512.

http://en.wikipedia.org/wiki/AVX-512#New_instructions_in_AVX-512_foundation

Is there any insight onto how this will effect Julia?

Is this something that Julia code can leverage? or is this something that is going to be compiler dependent?

[Stefan Karpinski]

We this kind of thing for free as soon as LLVM adds support. This is one of the great things about LLVM.

[Viral Shah]

This is exactly the kind of thing Arch has been pushing. I think with Intel taking LLVM seriously, we can expect to get these enhancements even sooner than in the past.

As Stefan said - Julia doesn't have to do too much to benefit. Also, all our various libraries will start using these things even sooner, which we will also benefit from.

-viral

[Erik Schnetter]

OpenBLAS may need re-tuning. I'm sure they are aware of this.

openlibm will require fundamental changes to see the benefits of AVX-512. Maybe there is an open-source Intel library that we could use instead of openlibm in this case?

-erik

--
Erik Schnetter <schn...@gmail.com>
http://www.perimeterinstitute.ca/personal/eschnetter/

My email is as private as my paper mail. I therefore support encrypting
and signing email messages. Get my PGP key from https://sks-keyservers.net.

[Viral Shah]

I believe they have been talking at Intel about open sourcing these, but not yet. We can easily incorporate any such projects into openlibm, and have a lower-accuracy mode, speed-accuracy tradeoffs etc.

-viral

[Eric Diaz]

I just saw this ...

from GCC GNU

https://gcc.gnu.org/wiki/cauldron2014?action=AttachFile&do=get&target=Cauldron14_AVX-512_Vector_ISA_Kirill_Yukhin_20140711.pdf

and from LLVM

http://llvm.org/devmtg/2013-11/slides/Demikhovsky-Poster.pdf

Some of the math stuff looks pretty amazing! For example,

-- AVX-512 exponential and reciprocal instructions contain more accurate approximate reciprocal instructions than those in the AVX-512 foundation; relative error is at most 2^-28. They also contain two new exponential functions that have a relative error of at most 2^-23. 

That's some pretty small relative error!

[Steven G. Johnson]

On Thursday, February 5, 2015 at 4:26:47 PM UTC-8, Eric Diaz wrote:

Some of the math stuff looks pretty amazing! For example,

-- AVX-512 exponential and reciprocal instructions contain more accurate approximate reciprocal instructions than those in the AVX-512 foundation; relative error is at most 2^-28. They also contain two new exponential functions that have a relative error of at most 2^-23. 

That's some pretty small relative error!

It's not small in double precision, where epsilon is 2^-52, so these instructions are probably more for graphics and gaming than for scientific computing.

[Erik Schnetter]

Yes -- these instructions are good for single precision. Some of these can also be used as starting values for iterative algorithms in double precision; if you e.g. use a Newton method that doubles the number of correct digits with each iteration, then one additional iteration will suffice for a double-precision result.

-erik