On Thu, 26 Jul 2012, Ronald S. Bultje wrote: > Hi, > > On Thu, Jul 26, 2012 at 3:51 PM, Loren Merritt <lor...@u.washington.edu> > wrote: > > 14% faster on penryn, 2% on sandybridge, 9% on bulldozer > > --- > > libavfilter/vf_hqdn3d.c | 157 > > +++++++++++++++------------------------------- > > 1 files changed, 51 insertions(+), 106 deletions(-) > > Looks good. > > I am going to ask a very stupid question: why is this faster? I see a > lot of simplification, which is good, but I'm not quite sure which > part actually has a clear speed impact.
Old code's sline_offs and dline_offs confused gcc into incrementing the src and dst pointers rather than using x as an index reg. Old code did horizontal(x), vertical(x), temporal(x). There's a dependency chain between those 3 filters, so you need to interleave multiple loop iterations to get maximum throughput. OOE might theoretically handle that, but doesn't do so perfectly on the CPUs I tested. New code does vertical(x), horizontal(x+1), temporal(x); which requires less OOE. --Loren Merritt _______________________________________________ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel