Hi,
I compiled the CMT plugins with the g++ options -msse2 and -march=i686
and the freeverb denormals problem (seen with ardour on a dual Xeon
machine) seems to be _almost_ entirely fixed. Also, I don't see any
weird problems with the ardour host (yet :). But, I still see the DSP
load increasing from around 5% to about 15%, although now it is _much_
farther along in the tails of the reverb than before and I get _zero_
xruns. [before, the DSP shot up to 60-90% with tons of xruns and
usually the zombification of ardour].
Much cooler.
--Neil
PS: I also am using the new undenormalize inline function described in
an earlier message.
On Sat, Jan 22, 2005 at 03:17:28PM +0100, Christian Schoenebeck wrote:
>> I think for such an important piece like freeverb it might be worth
to add
>> architecture dependant assembly optimizations. E.g. with SSE(1) you
simply
>> have to change the SSE control register to get rid of the denormal
problem.
>> With that no denormal fixes for individual sample points are necessary
>> anymore which definitely pays off in performance.
Agreed, but I'm not sure how doing that will affect the host - ie does
-mfpumath=SSE (or whatever the option is) always leave the FPU stack in a
sane state when control goes back to the host? I dont think it would
normally return the FPU to 387 mode, which might do bad things to the
host occasionally.
OTOH it might be fine, I dont understand the interaction with the 387 and
SSE.
- Steve
Received on Sun Jan 23 00:15:08 2005
This archive was generated by hypermail 2.1.8 : Sun Jan 23 2005 - 00:15:08 EET