[LAD] Denormals / subnormals (again)
fons at linuxaudio.org
Tue Jan 3 02:17:18 UTC 2012
On Mon, Jan 02, 2012 at 08:56:49PM -0500, Paul Davis wrote:
> On Mon, Jan 2, 2012 at 8:51 PM, Fons Adriaensen <fons at linuxaudio.org> wrote:
> > I haven't seen a single case where it wasn't easy to avoid denormals.
> > All it takes is some small offsets at strategic places in your code.
> > Usually that amounts to very few, and they are easy to spot if you
> > understand the code.
> hosts don't have the luxury of assuming that the author of a plugin
> smart enough to do Foo is also knowledgeable enough to take care of
> denormals. Lets face it - they are an artifact of the Intel
> architecture (powerpc and alphas never had this issue), and its
> reasonable to a host to assume that a plugin doesn't do the right
> thing. this would be true even if all but 1 plugin did, because you
> can pretty much guarantee that people will try to use the plugin and
> when it screws up the host, they will blame the host ...
> but sure. i mean even in a program as large as ardour, if DC bias is
> used to handle denormals, there's only one place where we add it.
I don't agree with blaming the Intel architecture. Denormals are
just one tiny aspect of a much wider issue which is numerical
precision and stability, and that affects all architectures.
Anyone writing numerical or DSP code should be aware of this
and be able to analyse the algorithm. You don't produce good
numerical or DSP code by blindly implementing textbook equations.
Vor uns liegt ein weites Tal, die Sonne scheint - ein Glitzerstrahl.
More information about the Linux-audio-dev