>>>>> "M" == Martin Hairer <[EMAIL PROTECTED]> writes:
M> There was some discussion on this list about interfacing lame
M> with Apple's G4 optimized FFT routines. Surprisingly, it seems
M> that lame spends basically no time computing FFT's on my G4
M> machine, but mostly square roots... At the lowest quality
M> settings, the percentages spent in the various routines
M> according to Shikari are as follows:
Probabry so.
The biggest problem on "NOT INTEL" machine is that IEEE_754 related hack
is not enabled by default.
These hacish code significantly speed up the code in quantize_xrpow(),
log(), log10(), and so on.
If you have a some "new" machine and want to fasten the LAME on it,
try to enable these hacks.
Define "TAKEHIRO_IEEE754_HACK" symbol and you can enable them.
If you get faster and proper result, pls report it to the list with
the your system descripton. We will update the configure script to
enable these hacks by default.
>>10.0% count_bits
>>3.7% ix_max
They are already SIMDed on ix86 machine with MMX.
I think at least ix_max() is easy to SIMDed.
>>3.8% sqrt
You may speed up the sqrt() by changing sqrt(x*sqrt(x)) to some polymorph.
Probabry this mail will help you.
http://sourceforge.net/mailarchive/forum.php?thread_id=1209312&forum_id=5500
--
Takehiro TOMINAGA // may the source be with you!
_______________________________________________
mp3encoder mailing list
[EMAIL PROTECTED]
http://minnie.tuhs.org/mailman/listinfo/mp3encoder
