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

Jed Brown jed at jedbrown.org
Wed Apr 30 16:51:20 UTC 2014


Brian Van Straalen <bvstraalen at lbl.gov> writes:

> I didn’t say we needed BLAS.  I specifically said BLAS and VSIPL are not useful for us.
>
> but like HPL, there is a large impact in how you write your benchmark if something like BLAS is around.
>
> In our case Jed is advocating SIMD operations.  
>
> I was trying to abstract vectorizing to help us converge on a simple portable code base. 

My code is portable and will automatically default to the (vanilla C99)
reference implementation if nothing else is available, otherwise it
default to the tuned version.

At the moment, we are dealing with two modes of extensibility:

  (a) different operators/algorithmic variants, and
  (b) tuned implementations.

We need both variables at this point in the design stage, and to make
(a) easier, I want to share as much of (b) as possible.  But eventually
(a) will no longer be variable, at which point sharing will no longer be
useful.  These are not deep code changes.
-------------- 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/20140430/26315ba6/attachment.bin>


More information about the HPGMG-Forum mailing list