Re: pb with -ffast-math

[Alain Baeckeroot]

Le 17/04/2009 à 10:28, Jaroslav Hajek a écrit :
> 
> Hmm, this looks really too complicated to vectorize. I guess a C++
> computation "kernel" taking care just of this operation is the way to
> go. Since there are no func calls, it should be fairly straightforward
> to write, using essentially the same syntax.
yes, Octave is nice :-)

> Are all the intermediate T1...T20 results really needed? They sound
> like temporaries to me.
> 
Part of them are results. So far we kept everything, to check the 
results were identical to the one given by the previous tool.
I'll try to eliminate the unneeded ones.

I have tested 3.1.55 compiled with "-O3 -march=native -funroll-loops"
it seems 10-20% faster.

we are also considering to add one dimension to all our variables,
 corresponding to differents independant input params, in order
to vectorize wrt this dimension, but i'm not sure this is useful
for our end-users !

Alain