Re: [eigen] Help on solving a race condition

[ Thread Index | Date Index | More Archives ]

On 09/06/2012, at 1:18 AM, Gael Guennebaud wrote:

> On Fri, Jun 8, 2012 at 5:09 PM, john skaller
> <skaller@xxxxxxxxxxxxxxxxxxxxx> wrote:
>> I may be missing the point here .. but why can't the cache sizes be
>> compile time constants? You can check once at startup the user supplied
>> values are correct and generate them in a configuration step.
> We don't have configuration step,

I know, but you could have :)

I mean we're talking about a #include file saying

#define L1SIZE ...
#define L2SIZE ..

so a reasonable guess is probably good enough.

> and how would you distribute a
> binary based on Eigen?

If the cache sizes disagreed the results may be suboptimal.
But it's hard to say much anyhow, if your system is virtualised
(like my Slicehost/Rackspace machine) or you're just running
more than one process, or you have a Bulldozer (2 cores but
only one FPU). How much performance are you losing NOT
using an available GPU?

If you just fixed the nominal cache size to some popular
values just how bad would it be?

john skaller

Mail converted by MHonArc 2.6.19+