Re: [eigen] some problems with building with AltiVec enabled

[jacob@xxxxxxxxxxxxxxx]

Quoting Gael Guennebaud <gael.guennebaud@xxxxxxxxx>:

> patch committed, thanks a lot.
>
> On Fri, Aug 22, 2008 at 12:37 PM, Konstantinos Margaritis
> <markos@xxxxxxxx> wrote:
> > Ok, I have an update of the AltiVec packetmath.h, seems ei_predux* methods
> > were buggy, so I fixed them. Plus, ei_pdiv, ei_pmul were also problematic, I
> > had to invent a new algorithm for ei_pmul (integer case) as AltiVec does not
> > have an instruction for that and most algorithms on that are about unsigned
> > ints (Eigen2 uses signed 32-bit ints). ei_pdiv was disabled for now, but I
> > intend to look at it again closer. Right now many tests pass, except these:
> >
> > * cholesky (some times it fails, 3/10 failures)
> > * geometry (3/10 failures)
> > * inverse (1/10 failures)
> > * sum (2/10 failures)
>
> I have the same here. The failures are likely due to numerical
> precision issues, and then just noise though we never checked that
> carefully.

I also have numerical stability issues. This is getting annoying  
because determining if a test passes or not often requires to run it  
several times (each time with 10 repeats) to see if failure occurs  
everytime. This is really unfortunate as it could make us ignore a  
real error that would be only occasionally triggerred.

When you say

> > intend to look at it again closer. Right now many tests pass, except these:
> >
> > * cholesky (some times it fails, 3/10 failures)
> > * geometry (3/10 failures)
> > * inverse (1/10 failures)
> > * sum (2/10 failures)

I hope that you mean that _sometimes_ they fail. Retrying, you should  
still be able to get 10 repeats working in a row. If it fails  
everytime you launch the default 10 repeats, then something is  
definitely wrong.

Cheers,
Benoit


----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.