27 Mar
2013
27 Mar
'13
12:48 p.m.
On Sun, Mar 24, 2013 at 05:18:40PM +0100, Raphaël BOLLEN wrote:
Thanks for looking into this, I can't check for
NaNs or Infs right
now but I've tried with an output from JAAA and have the same
result. Anyways the problem is solved with denormals protection as
suggested by Herman.
The code was supposed to be denormal-protected without those
patches, and I've found out why it isn't.
cmeterdsp.cc, lines 64,65 should be:
zl += _w1 * (*pl++ - zl) + 1e-15f;
zr += _w1 * (*pr++ - zr) + 1e-15f;
The current code uses 1e-20f as the offset. This is more than
enough to avoid denormals in the lowpass filters, but in the
following lines zl and zr are squared, and 1e-40f definitely
is denormal.
The same code is also used in jkmeter and jmeters. All will
be updated.
Ciao,
--
FA
A world of exhaustive, reliable metadata would be an utopia.
It's also a pipe-dream, founded on self-delusion, nerd hubris
and hysterically inflated market opportunities. (Cory Doctorow)