Subject: [linux-audio-dev] fp underflow
From: Juhana Sadeharju (kouhia_AT_nic.funet.fi)
Date: pe loka 08 1999 - 11:35:26 EDT
>I think you have to resort to adding DC, adding noise, testing for
>small values, that sort of thing.
Hello. I tested these ideas with my GVerb. Specially a reverb decays signals
toward zero but cutting the low amplitude tail away did decrease the
rate of exceptions only less than 5%.
Adding a DC feels like a bad idea but adding a DC at 24'th bit level
helped.
More I like the idea of adding noise. I added 24'th bit level noise
to input signal. The slowdown when underflow exception is on is
(1) 1.7 times without noise,
(2) 1.07 times with noise added to input.
The above estimates are based on the observation that a multiply with
underflow is 8.5 times slower than multiply without underflow.
I think adding noise can be done only the main inputs of the entire flow
network. Perhaps quantizer plug-in and already 24-bit signals are exceptions.
The effect of adding noise and DC were exactly the same.
Yours,
Juhana
This archive was generated by hypermail 2b28 : pe maalis 10 2000 - 07:27:13 EST