[Ffmpeg-devel] swscaler integration

[Erik Slagter]

On Fri, 2006-03-17 at 05:29 -0500, Dave Dodge wrote:
> On Thu, Mar 16, 2006 at 11:34:19AM -0500, Rich Felker wrote:
> > On Thu, Mar 16, 2006 at 12:58:14PM +0100, Erik Slagter wrote:
> > > I never understood why no unix version has a ioctl to set the pipe
> > > buffer size.
> > 
> > I guess then the kernel implementation would be a lot more complex
> > (not just using a single page) and there would have to be a mechanism
> > to prevent unbounded memory usage by pipe buffers.
> 
> It's been tried in Linux:
> 
>   http://www.ussg.iu.edu/hypermail/linux/kernel/0110.2/0693.html
> 
> The basic result was that on SMP systems they were able to improve
> performance on some benchmarks up by over 350%.  However on non-SMP
> increasing the pipe buffers degraded things by up to 30%.
> 
> This test was done several years ago on a 2.4.x kernel.  I seem to
> recall Linus rewriting the pipe implementation in the last year or two
> so the results might not apply any more.

I have been studying the linux pipe code some time ago, and I must say
it is very simple/basic (which is good, imho) and therefore it's
probably quite fast. Still for some applications this fixed buffer
length simply won't do. Maybe something for the shell developers to
"fix" by adding a second pipe type (which is actually a socketpair).
IIRC ksh had something like coprocesses which uses socketpairs.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/x-pkcs7-signature
Size: 2771 bytes
Desc: not available
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20060317/2b948db1/attachment.bin>