[Ffmpeg-devel] MMX/MMX2 and SSE optimizations for H.264 decoding

[Måns Rullgård]

Rich Felker <dalias at aerifal.cx> writes:

> On Thu, Sep 22, 2005 at 10:47:51AM -0700, Loren Merritt wrote:
>> On Thu, 22 Sep 2005, Romain Dolbeau wrote:
>> >Loren Merritt wrote:
>> >
>> >>we have:
>> >>put_h264_qpel_16_mc*_altivec
>> >>put_h264_chroma_mc8_altivec
>> >>
>> >>we do not have equivalents of:
>> >>put_h264_qpel_8_mc*_mmx2
>> >>put_h264_qpel_4_mc*_mmx2
>> >>h264_*_loop_filter_*_mmx2
>> >>ff_h264_idct_add_mmx2
>> >
>> >The additional question is, are we going to be bitten by
>> >Amdahl's Law ? i.e. are these important enought during
>> >the decoding process to justify optimizing them ?
>> 
>> idct is not so important (iirc 3% total speedup when I mmxed it).
>> qpel_4 matters only on videos that use 4x4 motion partitions.
>> qpel_8 and loop_filter are important.
>
> IMO loop filter is irrelevant because everyone who cares about
> performance disables it. :)

I though you cared about image quality.  Disabling the loop filter
degrades the image.

-- 
M?ns Rullg?rd
mru at inprovide.com