[FFmpeg-devel] ICC version checks

[Carl Eugen Hoyos]

Frank Barchard <fbarchard <at> google.com> writes:

> > > __attribute__((cold))
> >
> > This is (together with alloc_size) Intel issue 540761.
> >
> > > __attribute__((flatten))
> >
> > Intel issue 554092.

[...]

> Intel say the first 2 are 'mostly complete' but not released.

Good news, thank you for asking!
Allow me to repeat (I wrote the same in a private mail) that I believe the
vc1_mc_4mv_chroma() bug is significantly more important;-)

[edited]
 
> Thilo thinks its O3 related.

It is (It disappears for -O2 and -O3 -parallel).

> As a work around, I'd suggest an ifdef and
> #*pragma optimize* ( "", off )
> around the code you suspect.  That would also help Intel track down the bug.

It would not help significantly, I already posted that the problem is
decode_var_block_data(). Nobody wanted to look into the assembly of
vc1_mc_4mv_chroma(), so I will not post asm files if not explicitly asked to.

> They ask if you could file a bug for the ALS bug.

The reason I wanted to wait for our 0.6 release is that it was extraordinarily
difficult to report the vc1 bug, and it will be even more difficult if svn is
involved. I will try to put a alsdec-only version of FFmpeg together this
weekend.

Carl Eugen