13 Mar
2007
13 Mar
'07
11:36 a.m.
On Tue, 2007-03-13 at 10:08 -0500, Dave Phillips wrote:
So, if anyone has anything to add to Chris's
assessment, I'd like to
know about it. If the problem is indeed related to denormals, is there
a
way to fix it ? Comments and suggestions are most welcome.
i made a few experiences regarding to denormals in Aldrin/libzzub.
numerous plugins had their own fixes for denormalization (and i know
that CAPS has a fix too, which is to add a -250dB square wave to avoid
the value to go off the scale). AMD cpus do not have that problem most
of the time, so the problem relates mostly to Intel cpus.
If all plugins and the host use SSE fp arithmetic, the host can turn off
denormalization through SSE intrinsics for each render step.
We did exactly that. I enabled the builds for SSE, removed any
denormal-fix related code and just disabled denormals. it makes the code
a lot cleaner and faster, however the downside is that non-SSE cpus will
not benefit from that advantage.
--
Leonard Ritter
-- Freelance Art & Logic
-- http://www.leonard-ritter.com