[Ffmpeg-devel] Re: [PATCH] fix snow x86 SIMD
Martin von Gagern
Wed Aug 9 22:11:34 CEST 2006
The previous two patches submitted by Luca Barbato were originally from
me. Now I subscribed to the list myself, to accept the blame personally.
Michael Niedermayer wrote:
> sorry, no, every patch you submit gets dirtier then the previous, either
> cleanly change the code or leave it
OK, I tried another approach, and I believe it is cleaner from the
register allocation perspective. It leaves all those register names to
the compiler, by declaring them as "=&r" output parameters associated to
otherwise unused variables. With optimization turned on, the values
stored in those registers are discarded after the asm block and don't
get written to the variables' memory location. With -fomit-frame-pointer
alongside -fPIC, this compiles well enough, as the compiler uses %ebp.
> additionally, id suggest that you run the regression tests before sumitting
> patches, (i think ill ignore every asm related PIC fix if the author doesnt
> explicitly say that the regression tests passed)
Now I'm compiling ffmpeg the official way, without any Gentoo specific
overhead. I tried running the regression tests. Unfortunately the flac
codec fails for my modified working copy as well as for current clean
r5969 and for r5948. I tried that last one because I found a mail where
someone explicitly had regression tests working with that revision. The
program crashes during that test. I'll investigate it further and start
an independent thread on this once I know more.
In the meantime, I ran the snow and snowll tests. Make failed for both,
but the diff listed four matching lines each time, so I'd consider that
a pass for the snow codec. And unless snowdsp_mmx.c is used for
something not related to snow, I'd guess everything else works as it did
before - or didn't in my case.
Martin von Gagern
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
More information about the ffmpeg-devel