[HPGMG Forum] Do we want the benchmark to go into intrinsics?
jed at jedbrown.org
Tue Apr 29 17:49:50 UTC 2014
Sam Williams <swwilliams at lbl.gov> writes:
> There's also a deeper question on how you evaluate/rank.
> I was thinking of requiring submitters to submit 6 numbers at a given concurrency.
> - reference implementation for N, N/8, and N/64 DOF/process
> - full fury implementation for N, N/8, and N/64 DOF/process
> Ranking would be based on a linear combination of these (the first three could be weighted by 0)
This size-dependence is related to my "dynamic range" proposal. I'm not
committed to any specific weighting, but I think the ultimate metric
should have some way of rewarding architectures with good dynamic range.
For Edison, all 3 numbers you propose should be within a few percent.
So far, I have not been able to observe that on BG/Q.
If people care about the reference "non-optimized" implementation (e.g.,
if it is used for ranking), then they can have the compiler
pattern-match. I don't see how to adjudicate this in the long term.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 835 bytes
Desc: not available
More information about the HPGMG-Forum