Pentium performance tweaking, was Re: pd-0.21 on linux...

[Larry Troxler]

Charles Baker wrote:
> 
> Finally got pd-0.21 to compile on my linux platform. Mea culpa....
> I had installed 'egcs' , an experimental gcc 'improvement' (even higher
> degrees of
>  optimization than standard gcc available ...someone on the csound list
> reported a major increase in speed in csound compiled with this
> compiler)...

How much an increase, percentage-wise? 

As an aside, I notice an almost 100 % increase on Pentium-100 Linux when
I use a 128-sample DAC Block size instead of 64 samples! THis is based
on what "top" reports for CPU%, which is also consistent with what the
"profile" object reports. However, I don't yet know whether this is
perhaps 
due to the complex timeing interactions between the DAC size and the OSS
fragment size and the scheduler advance.

It think it would be usefull if someone else on similar hardware would
verify this. IIRC, it's the DACBLKSIZE parameter that I'm speaking of.


That much difference does seem a bit suspicious, but maybe it makes
sense due to the tiny first-level cache (each dsp performance loop
easily fits in to the cache and runs at breakneck speed). Or could there
be some inefficiency in what happens between DSP runs? I looked and
didn't immediately see any glaring errors, but ...

But regardless, 128 samples represents just under 3ms at 44100Hz, which
isn't too bad compared to MIDI latency. So it's not a bad deal :-)

Larry

--  Larry Troxler --  lt at westnet.com  --  Patterson, NY USA  --