Boost logo

Boost :

From: boost (boost_at_[hidden])
Date: 2002-05-06 04:58:16


Salut,

On Monday 06 May 2002 09:57, Matthias Troyer wrote:

> No, m*=0 does not work, as far as I know. At least in my experience
> (but it might be a broken implementation), if the uninitialized
> matrix contains NaN, multiplying the NaN with 0 is still NaN and not 0.

That's the correct behaviour for IEEE complying arithmetic.
Several years ago I read a comment on this in the BLAS
documentation, it's somewhere mentioned.
You have to be carefull with using BLAS with respect to this
behaviour, if you want to calculate A = B * C, since you have to
clear A first, if there's a possibility for A to have NaN entries,
since dgemm defines A = alpha*A + beta*A*B.

Best wishes,
Peter


Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk