[gmx-developers] double precision powerpc

[Erik Lindahl]

Hi,
>
>
> i noticed that kernels (in src/gmxlib/nonbonded/ 
> nb_kernel_ppc_altivec) optimized for powerpc have functions  
> accepting floats.  i don't see how this will work in double  
> precision.  the function prototype for the kernels is defined to  
> take 'real' as parameters (nb_kerneltypes.h) and 'real' is  
> typedef'ed to either float or double since my toolchain do not  
> support 'real'.

The configuration script will stop you from ever compiling double  
precision with altivec, so that can't happen.


> is someone already working to get this working? or is there another  
> way to get ppc_altivec working with doubles?

Nope, the VMX/Altivec hardware unit only supports single precision  
floating-point. Double precision Altivec is impossible until that  
changes.

Cheers,

Erik