Re: Polyhedron b'marks | Linux vs Windows

[Tim Prince <n8tm at aol dot com>]

On 6/3/2014 9:43 PM, Jerry DeLisle wrote:
> On 06/01/2014 03:57 PM, Evan Cooch wrote:
> > So, have a box running a new-ish, fairly high-end Intel multi-core chip
> > (which one doesn't really matter here). Decided to see if the OS
> ---- snip ----
>
> A few years ago I was investigating some of these benchmark issues.
>
> I do not remember all the details, but I did find that at least one of the
> common libraries was implementing some of the basic trig functions, such as
> sin(), with software only methods and was not using the floating point
> co-processor capability in most modern CPUs.
>
> I attempted to get a maintainer to update this code and use the available
> hardware.  I was met with a response more or less "Does it make a real
> difference in a real application?"  That is a legitimate question and I don't
> have an answer.
The built-in x87 math function support is at a disadvantage when called 
from SSE or AVX code, due to the necessity of passing operands and 
results by stack or memory reference.  It makes sense to use it only for 
real(10) now that x87 mode is seldom used for single or double precision.
The usual method for in-lining x87 math functions was via the 
mathinline.h which can be used outside of glibc, but normally isn't used 
with sse math.   The extra bugginess of mathinline.h in several 
prominent linux distros gave it a bad name.  I suppose the builtins 
scheme was intended to provide an alternative but came too late.
The big difference is in possible use of simd vector math functions such 
as Intel VML or svml library functions, which aren't intended for use 
with gfortran on Windows.  I assumed these weren't used with gfortran on 
linux either.  With Intel compilers, the omp simd directives request use 
of svml even when the compiler wouldn't otherwise call them (including 
combining simd and threading by omp parallel do simd).
I thought the newlib math functions were similar enough to the glibc 
ones; anyway, the OP didn't care to answer about which Windows 
implementation of gfortran or math library was under discussion.

--
Tim Prince