[Ffmpeg-devel] [PATCH] Snow mmx+sse2 asm optimizations

Robert Edele yartrebo
Fri Mar 10 01:14:24 CET 2006


On Thu, 2006-03-09 at 15:23 -0800, Loren Merritt wrote:
> On Thu, 9 Mar 2006, Robert Edele wrote:
> 
> > I've removed interleave_line_footer which is never invoked, as it is not
> > used by the asm cases and Michael's code is used for the C case.
> >
> > Michael, I believe that the patch is ready to be reviewed and hopefully
> > committed. There might be a small bug or two left, but the code looks
> > good to my eyes and has been tested on AMD-64, P4, and AMD-32 systems
> > with no issues (no crashes, and md5sums all agree). Regression tests on
> > my machine all pass (C, mmx, sse2).
> >
> > If there are any more remaining issues with the code, please let me
> > know. Thanks for your patience.
> 
> >+ if (!(b_h & 1))
> >+     inner_add_yblock_bw_8_obmc_16_bh_even_sse2(obmc, obmc_stride, 
> >+         block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
> >+ else
> >+     inner_add_yblock_bw_8_obmc_16_bh_even_mmx(obmc, obmc_stride, 
> >+         block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
> 
> inner_add_yblock_bw_8_obmc_16_bh_even_mmx is called when bh is odd?
> 
> --Loren Merritt

That must have been an oversight by Oded (ods15) while helping me out in
porting the code out of snow.c and renaming that function after reading
the corresponding sse2 function (which does require even b_h). The mmx
version handles either odd or even just fine and it was never written
with the word 'even' in it. On the plus side, he is the one that thought
of using the mmx as a fallback should b_h be odd instead of falling back
on the C version.

I've attached the patch with the naming fixed. Thank you for bringing
this to my attention.

Sincerely,
Robert Edele
-------------- next part --------------
A non-text attachment was scrubbed...
Name: snow_mmx.patch
Type: text/x-patch
Size: 93586 bytes
Desc: not available
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20060309/080fa7de/attachment.bin>



More information about the ffmpeg-devel mailing list