[LAD] Denormals / subnormals (again)
lists at carlh.net
Tue Jan 3 01:22:11 UTC 2012
I have been looking into ye olde denormal problem a little, lately.
Particularly with respect to Ardour and plugins. I've assembled what I
believe to be a coherent statement of what is going on, but I'd very much
appreciate any corrections and clarifications that anyone can offer.
Here's how it seems to me:
If you compile your code (e.g. a plugin) without -msse and -mfpmath=sse on
the GCC command line you get *no* protection from denormals from the CPU.
If they occur in your code, they will be very much slower than normal
floating point numbers (~49 times slower on my Core 2 Duo, ~7 times slower
on a Core i3). As far as I can see, it does not matter that Ardour has
been built with those flags: if a plugin has not, you have no protection.
If you compile your code with -msse and -mfpmath=sse, you have Ardour's
protection from denormals. If the user's CPU supports it, you get
there is no significant slowdown with denormals using this mode. However
CPU support is "some later processors with SSE2", according to Intel.
The problem, I guess, is that we cannot really distribute plugins with SSE
instructions, otherwise we do not support people with older CPUs. In this
case, I think the plugin code must avoid denormals, otherwise there will
be a significant performance hit if they arise.
I've been testing behaviour using a very dumb program which you can get
I've also been testing plugins using a primitive torture tester that you
can get from http://carlh.net/software/ or on github via
git at github.com:cth103/plugin-torture.git
More information about the Linux-audio-dev