Hello Wizards.

I'm sorry for posting such a general question on such a specific newsgroup, but i believe here are people who might know how to help.

The usual problem in sound processing using floating-point is denormalization. The bad thing is, it usually happens not to one value, but to thousands at once. Is there any way to deal with it without a huge penalty? It is not necessary to try to treat de-normals carefully, it is enough to zero them out. Is there an efficient way to implement this? I know there is within SIMD2 (Pentium4), but the code has to stay at least Pentium3/Athlon compatible.

Does the program start-up code set the FP control word? Are the FP exceptions enabled or disabled?

Thanks in advance,
-i.

"Ilya Minkov" <midiclub@8ung.at> wrote in message news:b9h8uf$187i$1@digitaldaemon.com...
> The usual problem in sound processing using floating-point is denormalization. The bad thing is, it usually happens not to one value, but to thousands at once. Is there any way to deal with it without a huge penalty? It is not necessary to try to treat de-normals carefully, it is enough to zero them out. Is there an efficient way to implement this? I know there is within SIMD2 (Pentium4), but the code has to stay at least Pentium3/Athlon compatible.

The easiest way is to switch to 80 bit reals. That gives you more room before you lose bits to denormalization.

> Does the program start-up code set the FP control word?

Yes.

> Are the FP
> exceptions enabled or disabled?

Disabled.