[Bug tree-optimization/55016] New: request for specific builtins for rcp and rsqrt

[Bug tree-optimization/55016] New: request for specific builtins for rcp and rsqrt

[vincenzo.innocente at cern dot ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55016

             Bug #: 55016
           Summary: request for specific builtins for rcp and rsqrt
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: vincenzo.innocente@cern.ch


There are cases where the use of approximate rcp and rsqrt suffice.

I wonder if it would be possible to introduce specific "generic" builtins for
"rcp" and "rsqrt" that produce the proper instruction depending on the target
architecture (see,avx etc) and eventually generate vector instruction in a loop

at the moment anything like this is target specific, inefficient and does not
vectorize!

#include <x86intrin.h>
float v0[1024];
float v1[1024];
inline
float rsqrtf( float x ) {
  return _mm_cvtss_f32( _mm_rsqrt_ss( _mm_set_ss( x ) ) );
}
void v() {
  for(int i=0; i!=1024; ++i)
    v0[i] = rsqrtf(v1[i]);
}

[Bug tree-optimization/55016] request for specific builtins for rcp and rsqrt

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55016

--- Comment #1 from Marc Glisse <glisse at gcc dot gnu.org> 2012-10-22 18:56:42 UTC ---
(In reply to comment #0)
> void v() {
>   for(int i=0; i!=1024; ++i)
>     v0[i] = rsqrtf(v1[i]);
> }

Doesn't writing
v0[i] = 1 / sqrtf(v1[i])
work with suitable fast-math flags? It still produces an extra iteration to
refine the result, do we want a -ffaster-math?

[Bug tree-optimization/55016] request for specific builtins for rcp and rsqrt

[vincenzo.innocente at cern dot ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55016

--- Comment #2 from vincenzo Innocente <vincenzo.innocente at cern dot ch> 2012-10-23 05:19:37 UTC ---
For the application I have in mind a global flag will such as -ffaster-math
will not be suitable
as it would affect also places where full "single precision" is still required.
I would like just to profit of the rcp and rsqrt instructions for cases where
their low "precision" is enough.

[Bug tree-optimization/55016] request for specific builtins for rcp and rsqrt

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55016

--- Comment #3 from Marc Glisse <glisse at gcc dot gnu.org> 2012-10-23 06:12:30 UTC ---
(In reply to comment #2)
> For the application I have in mind a global flag will such as -ffaster-math
> will not be suitable
> as it would affect also places where full "single precision" is still required.

Flags can be used per function, but that doesn't work very well indeed.

> I would like just to profit of the rcp and rsqrt instructions for cases where
> their low "precision" is enough.

Some math libraries (AIX for instance) provide a correctly rounded rsqrt, so it
would have to be called vaguely_estimate_rsqrt or something (still not to be
confused with the fast-math version that adds one refinement round to reach
almost full precision).