[Spice-devel] libjpeg performance

[Alexander Larsson]

On Wed, 2010-04-14 at 09:33 +0300, Uri Lublin wrote:
> On 04/12/2010 09:57 PM, Alexander Larsson wrote:
> > I did some simple testing of the new mjpeg encoder.
> > Showing the youtube "will it blend - ipad" video i got quite better
> > compression (24k per frame average as opposed to 35k before), but the
> > code used a bit more cpu (9.8 msec per frame where it was 4.2 before). I
> > made a simple change in the libjpeg code to make it a bit faster, but we
> > might still be able to tweak this a bit in favour of performance rather
> > than compression if thats what we want.
> >
> 
> Hi Alex,
> 
> Nice.
> 
> In  client/mjpeg_decoder.cpp -> decode_data(), there is a code that memmove data 
> to the beginning of the buffer.
> 
> Do you think we can gain better performance by actually implementing _cinfo.src 
> functions ? This way we may be able to avoid (some) memory copying, e.g. by 
> using a circular-buffer (and modulo calculation) or maybe even with no internal 
> buffer.

In current use we actually always pass an entire frame of data to
decode_data() each time. I think rather than doing some fancy buffer
work we could just special case that case and avoid copying data at all.
i.e. if there is no existing data, just set _jsrc.next_input_byte to
point to to data, and if not everything is copied (unlikely given the
above), just copy the remainder into _data afterwards.

-- 
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
 Alexander Larsson                                            Red Hat, Inc 
       alexl at redhat.com            alexander.larsson at gmail.com 
He's a gun-slinging coffee-fuelled inventor from the Mississippi delta. She's 
a bloodthirsty insomniac doctor from a different time and place. They fight 
crime!