On 05/28/2011 01:37 AM, Jörn Nettingsmeier
wrote:
On 05/27/2011 10:08 PM, Fons Adriaensen wrote:
I've
got a strange problem with MFV. *Something* is poisoning
the signal graph with Nans or Infs. One visible result is that
e.g. the W channel of one of the jkmeters shows only peaks and
no RMS values (this is to be expected if the filter state gets
a Nan - once a Nan always a Nan). The other indication is that
adding an Ambdec produces a never ending xrun storm even if the
effective CPU load remains below 10%.
Have you seen anything similar ?
not in this session. sc4 can be funny at times, but i haven't had any
issues with it for quite a while. i don't have access to the session
atm, will ake a look when i return home on sunday. meanwhile, if you
want to pinpoint the problem, try muting all channels and then unmute
one by one.
do the nans crop up immediately, or only after playing some audio?
i just tested the session again, and i don't see any NaNs or INFs. what
i did notice, though, is that i'm burning a ridiculous amount of cpu for
this tiny little arrangement - i had tried with 64 frames initially, and
it maxes out immediately (on an 2.6 ghz i5 quad-core). safe way to run
it is at -p 1024 -n2.
Turned out to be jconvolver. Added some Nan testing code to zita-convolver
and the problem disappeared, also without the extra code. Could be I
accidentally installed a version compiled on another machine. But then
it's a miracle non of this showed up earlier...