Re: Denormalization
Re: Denormalization
- Subject: Re: Denormalization
- From: Jeff DuMonthier <email@hidden>
- Date: Wed, 11 Dec 2002 21:52:13 -0500
Steinberg have noted on their dev list, that they usually do this
virtually in any dsp application, so I think it has been tested
successfully over the past years. Some other developers stated,
they even don't restore the registers (okay, I think these were
Wintel guys)...
The performance benefit on tight loops can in some cases be
probably around 10%, maybe more if pipeline stalls can additionally
be avoided (or if you use a branching solution to prevent
denormalization :-).
I've seen the CPU usage of the scalar version of my VST plugin jump
by a factor of two when it hit denormals. I think the factor was
close to 8 in some of my early AltiVec experiments when the data
degenerated to denormals. I always set the non-Java bit for AltiVec
code.
Unfortunately, I still have to compile a pre OS X version with MrCpp,
and the gcc code won't work for that. By the way, I cannot restore
the previous state of the bit for AltiVec code built with MrCpp
because vec_mfvscr is missing from the MPW libraries. Since I have
no way of knowing whether or not it was set to begin with I just set
it and leave it.
--
-Jeff DuMonthier
_______________________________________________
coreaudio-api mailing list | email@hidden
Help/Unsubscribe/Archives:
http://www.lists.apple.com/mailman/listinfo/coreaudio-api
Do not post admin requests to the list. They will be ignored.