[HPGMG Forum] Do we want the benchmark to go into intrinsics?

Jed Brown jed at jedbrown.org
Tue Apr 29 17:32:18 UTC 2014


Sam Williams <swwilliams at lbl.gov> writes:

> Bad compilers should be shamed.  Having it autodetect __bgq__ (or
> whetever) and run the best manually optimized implementation by
> defaults hides deficiencies.

I agree and it would be interesting to require that official results use
portable code.  However, as we discussed at the meeting, many of these
vendors have their own compilers so they can just pattern match on the
benchmark code and insert the hand-optimized code.  (This is more
expensive than optimizing the benchmark code directly, but less
expensive than making the compiler work well on user code.)

Anyway, that is a legislation problem that I would like to defer for
further community discussion.  At this moment, I want to be able to say
things about the way an algorithm exercises hardware.  FWIW, performance
is still terrible even with the QPX intrinsics, though less terrible
than before.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 835 bytes
Desc: not available
URL: <https://hpgmg.org/lists/archives/hpgmg-forum/attachments/20140429/7a0c62f5/attachment.bin>


More information about the HPGMG-Forum mailing list