opus - Mar 2015 - [RFC PATCHv2] Intrinsics/RTCD related fixes. Mostly x86.

Nit: in dual_inner_prod_sse, why not do both horizontal sums at the same 
time? As in:

     xsum1 = _mm_add_ps(_mm_movelh_ps(xsum1, xsum2), 
_mm_movehl_ps(xsum2, xsum1));
     xsum1 = _mm_add_ps(xsum1, _mm_shuffle_ps(xsum1, xsum1, 0xf5));
     _mm_store_ss(xy1, xsum1);
     _mm_store_ss(xy2, _mm_movehl_ps(xsum1, xsum1));

--John

This causes regression for ARMv7 Neon intrinsics.. I'm fixing this
now.. will push out patch this week end.

Regards,
Vish


On 12 March 2015 at 15:24, John Ridges <jridges at masque.com>
wrote:
> Nit: in dual_inner_prod_sse, why not do both horizontal sums at the same
> time? As in:
>
>      xsum1 = _mm_add_ps(_mm_movelh_ps(xsum1, xsum2),
> _mm_movehl_ps(xsum2, xsum1));
>      xsum1 = _mm_add_ps(xsum1, _mm_shuffle_ps(xsum1, xsum1, 0xf5));
>      _mm_store_ss(xy1, xsum1);
>      _mm_store_ss(xy2, _mm_movehl_ps(xsum1, xsum1));
>
> --John
>
> _______________________________________________
> opus mailing list
> opus at xiph.org
> http://lists.xiph.org/mailman/listinfo/opus

The patch only takes the existing Opus SSE code and moves it (out of header
files, into C files) in order to make it possible to use it with RTCD.  There
aren?t any changes to the code itself.

Improvements to the SSE code are very likely possible, but I think they should
be contributed as a separate patch. This patch is too big already.

On Mar 12, 2015, at 4:24 PM, John Ridges <jridges at masque.com> wrote:
> Nit: in dual_inner_prod_sse, why not do both horizontal sums at the same 
> time? As in:
> 
>     xsum1 = _mm_add_ps(_mm_movelh_ps(xsum1, xsum2), 
> _mm_movehl_ps(xsum2, xsum1));
>     xsum1 = _mm_add_ps(xsum1, _mm_shuffle_ps(xsum1, xsum1, 0xf5));
>     _mm_store_ss(xy1, xsum1);
>     _mm_store_ss(xy2, _mm_movehl_ps(xsum1, xsum1));
> 
> --John
> 
> _______________________________________________
> opus mailing list
> opus at xiph.org
> http://lists.xiph.org/mailman/listinfo/opus