On Thu, Nov 26, 2015 at 11:15 AM, John Warburton wrote:
> On Thu, Nov 26, 2015 at 1:26 PM, Ganesh Ajjanagadde wrote:
>
>>> Does removing the L fix it?
>
>> Read up a bit and suspect it should, and more generally long double is
>> a can of worms on ppc with bugs on some compilers:
>> https://gcc.g
On Thu, Nov 26, 2015 at 1:26 PM, Ganesh Ajjanagadde wrote:
>> Does removing the L fix it?
> Read up a bit and suspect it should, and more generally long double is
> a can of worms on ppc with bugs on some compilers:
> https://gcc.gnu.org/wiki/Ieee128PowerPC, ABI in transition,
> https://llvm.org
On Thu, Nov 26, 2015 at 8:07 AM, Ganesh Ajjanagadde wrote:
> On Thu, Nov 26, 2015 at 5:23 AM, John Warburton
> wrote:
>> On Fri, Nov 13, 2015 at 4:42 PM, Ganesh Ajjanagadde
>> wrote:
>>>
>>> 4/4 is a "best effort" patch that maximizes accuracy on all platforms for
>>> avcodec/faandct. It result
On Thu, Nov 26, 2015 at 5:23 AM, John Warburton wrote:
> On Fri, Nov 13, 2015 at 4:42 PM, Ganesh Ajjanagadde
> wrote:
>>
>> 4/4 is a "best effort" patch that maximizes accuracy on all platforms for
>> avcodec/faandct. It results in concrete accuracy benefits on the "default"
>> x86-64
>> GNU/Lin
On Fri, Nov 13, 2015 at 4:42 PM, Ganesh Ajjanagadde
wrote:
>
> 4/4 is a "best effort" patch that maximizes accuracy on all platforms for
> avcodec/faandct. It results in concrete accuracy benefits on the "default"
> x86-64
> GNU/Linux platform.
>
> Patches tested with FATE. ppc untested.
Sorry t
1/4 uses predefined mathematics macros whenever possible to improve readability
and in some cases accuracy.
2/4 is a minor "pedantic" patch for something noticed while working on this
stuff.
3/4 is a more important patch, since the precomputed cosines on ppc are
terrible,
with accuracy off even