[x265] Performance hit with revision 341049cdcc5c

[Steve Borho]

On Thu, Aug 29, 2013 at 7:15 AM, Nikos Barkas <nikbar2004 at yahoo.com> wrote:

> I note a small but measurable performance hit with revision 341049cdcc5c
> on Windows, compared to earlier revisions. This is on a MinGW build.
>
> My standard encode test run normally takes approximately 57 seconds to
> complete and hits 1.74-1.75 fps. This revision takes approximately 64.4
> seconds, with performance at 1.55 fps.
>
> Looking at the commit, it seems like a blockcpy instruction was replaced
> with a quite a bit more complicated logic. It is probably expected that
> performance might be affected, especially if the method where this logic is
> used runs continuously, but I thought I'd bring this up because it seems
> like we have a bottleneck and it could be a good idea considering options
> for optimizing this part.
>

We're in the middle of replacing pre-computed (and thus stored in memory)
16 subpel (and perhaps weighted) reference planes with on-demand
generation.  In it's current state we are doing both pre-generation and
partial on-demand, so we're paying double price.

This should all get resolved over the next couple days.

-- 
Steve Borho
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.videolan.org/private/x265-devel/attachments/20130829/3306f1e3/attachment.html>