[FFmpeg-devel] Patch: Inline asm fixes for Intel compiler on Windows

[Matt Oliver]

>
> This is confusing, I thought the speedup was from removing the lea?
> Changing the type would be on top.
> So you mean that after the first patch, changing the type causes a ca. 4%
> slowdown?


Yeh, both variants are in fact faster than what is there currently as the
speedup is from replacing the lea. But the version with uint8 is not as
fast as the one provided in the last patch.


> That should be wrong. All the ifs should be using the _branchy variant
> which does not use the asm code at all.
>

Well if it should be then its not. All the functions in vp5.c use
vp56_rac_get_prob not the branchy version. Look at vp5_parse_vector_models
for example. It appears vp8 and vp9 use the branchy version which is why i
singled out vp5 as one that may benefit. However it appears that vp5.c may
be in need of some updates to use the branchy version.

If these should be using the branchy version then there is really no reason
to use uint8. So the current patch would appear to be the way to go.