Re: [PATCH][ppc] merge in powerpc version of slowpow.c

[Andreas Jaeger <aj at suse dot com>]

On 01/03/2013 10:20 AM, Siddhesh Poyarekar wrote:
> Hi,
>
> The powerpc version of slowpow.c adds an extra pass with the long
> double functions, which reduces the number of values that fall back
> into the slow mpa phase.  This should be useful for other

Do you have some examples that still fall in the slow mpa phase?

> architectures as well if they have support for a larger long double.
> Attached patch merges in the powerpc slowpow.c logic into the globally
> used code and removes the former.  I have verified that the test suite
> does not regress on powerpc 32 and 64-bit as well as on x86_64.
>
> The test cases in http://entropymine.com/imageworsener/slowpow/ are
> now useless since all of those values get computed correctly using
> long double and return in a flash.  The powerpc performance with this

So, for x86-64 this is a great improvement - thanks!


> change has improved by about 1%, which should be due to the favourable
> branch prediction, which is biased towards finding an accurate result
> using long double.
>
>
> Siddhesh
>
> 	* sysdeps/ieee754/dbl-64/slowpow.c (__slowpow)
> 	[!NO_LONG_DOUBLE]: Try computing pow using long double.
> 	* sysdeps/powerpc/powerpc32/power4/fpu/slowpow.c: Remove.
> 	* sysdeps/powerpc/powerpc64/power4/fpu/slowpow.c: Likewise.
>
> diff --git a/sysdeps/ieee754/dbl-64/slowpow.c b/sysdeps/ieee754/dbl-64/slowpow.c
> index c303eaa..eafd6e1 100644
> --- a/sysdeps/ieee754/dbl-64/slowpow.c
> +++ b/sysdeps/ieee754/dbl-64/slowpow.c
> @@ -47,13 +47,33 @@ double
>   SECTION
>   __slowpow(double x, double y, double z) {
>     double res,res1;
> +
> +  res = __halfulp (x, y);	/* halfulp() returns -10 or x^y             */
> +  if (res >= 0)
> +    return res;			/* if result was really computed by halfulp */
> +  /*  else, if result was not really computed by halfulp */
> +
> +#ifndef NO_LONG_DOUBLE
> +  long double ldw, ldz, ldpp;
> +  static const long double ldeps = 0x4.0p-96;

Is that constant correct for all the various long double formats we have?

> +
> +  /* Compute pow as long double, 106 bits */

The comment 106 bits is not correct and needs updating,

Andreas
> [...]

--
 Andreas Jaeger aj@{suse.com,opensuse.org} Twitter/Identica: jaegerandi
  SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
   GF: Jeff Hawn,Jennifer Guild,Felix Imendörffer,HRB16746 (AG Nürnberg)
    GPG fingerprint = 93A3 365E CE47 B889 DF7F  FED1 389A 563C C272 A126