[FFmpeg-devel] [PATCH] Optimization for add_8x8basis

[Loren Merritt]

On Sun, 13 May 2007, Zuxy Meng wrote:
> 2007/5/13, Guillaume POIRIER <poirierg at gmail.com>:
>> On May 13, 2007, at 3:56 , Zuxy Meng wrote:
>>
>> 3DNow! and SSSE3 provide instructions for packed mulplication with
>> rounding (pmulhrw and pmulhrsw respectively) which can be used to
>> replace pmulhw+paddw+psarw.
>>
>> Someone willing to test it on Core 2?
>>
>> What kind of sample would be a good test? MPEG-4 ASP, AVC, .... ?
>
> To be honest, I really don't have an exact idea of where it's used:-)
> The function add_8x8basis is called in mpegvideo.c, but even fulltest
> doesn't cover that particular code path.

add_8x8basis is used when encoding with qns enabled (in any mpeg-like 
format).

works for me (no change in output file).

benchmark on core2:
add_8x8basis_mmx=52, ssse3=38 cycles.

--Loren Merritt