single precision compile broken

[Axel]

On Mar 11, 10:34 pm, Nick <naro... at gmail.com> wrote:
> Axel and Teo,

nick,

> I'm interested in learning more about the state of the single
> precision compilation. If it works correctly,

no. it does not work. well, you can compile and some
inputs run (most regtests).

> the speed up is a factor of 4x - 8x (?) over double precision with no

why a speedup of 4x-8x???

> lost in accuracy I presume.

au contraire! a _lot_ of accuracy is lost.
that is the very nature of running in single precision.

>
> Some questions:
> 1. Does it work in parallel?

no.

> 2. Does it work with QS + OT?
in principle, yes.

> 3. Are the standard LDA & PBE functionals supported?

the code used is essentially identical to the DP
code, only the _dp kind is changed and the corresponding
single precision versions of BLAS/LAPACK are called.
...and then there are a few special subroutines.

> If so, can you summarize for us:
> 1. Compiler versions which work, plus optimization levels.
> 2. math libraries that work
> 3. FFTW2 should be straightforward in single precision

well, and i'd really like a pony for next christmas. ;-)

cheers,
   axel.


>
> On Jan 22, 8:24 am, Axel <akoh... at gmail.com> wrote:
>
> > On Jan 22, 12:15 pm, Teodoro Laino <teodor... at gmail.com> wrote:
>
> > > Hi Axel,
>
> > hi teo,
>
> > > anyway the intel compiler we have on our machines (10.0.025) is able
> > > to compile the today CVS snapshot.
>
> > it only happens withsingleprecision.
> > with the intel 10.0 compiler it crashes while compiling src/
> > glob_matching.F
> > turning off optimization works around it. my default optimization for
> > that
> > compile is very high. i'm doing this deliberately to be able to spot
> > problems.
> > this is the first time, it broke the compiler, though. it doesn't help
> > much with
> > the little trust that i have left in vectorization and automatic SIMD
> > optimizations
> > which are supposed to be _so_ much better with recent compilers.
>
> > > No internal compiler errors..
> > > As I said it's a 10.0.025 on an x86-64.
>
> > same here.
>
> > cheers,
> >    axel.
>
> > > Ciao,
> > > Teo