Subject: Re: [linux-audio-dev] Athlon64 .. performance boost ?
From: Juan Linietsky (coding_AT_reduz.com.ar)
Date: Thu Oct 16 2003 - 22:50:04 EEST
> Or on the other hand if one could use 64bit integer processing
> substituting it to floating point processing.
> Juan L. always tells me about the great speed you can achieve writing
> audio apps that work with integers but I always
> felt that 32bit offers too little headroom (imagine mixing hundreds of
> voices with individual volumes without loosing
> the original sample quality and providing some headroom for FX
> processing etc.
There is barely any difference between integers and floats for math
nowadays, if there is any, except that some operations can be optimized
using bit operations and/or fixed point, like modulus, division, flooring,
etc. The prime example on when your DSP algorithm becomes turtle
is when you need to convert from float to integer, so you can get a
value from a buffer. That alone can take a lot more time than the
rest of your dsp code. floor() fmod() are also examples of complete
slowdown (and then are many times necesary, for example, in something
like resampling).
But again, regarding to float->int conversion, As far as i know it is
specifically a poblem with x86 processors? I know there is one instruction
that does this, but it still takes several cycles.
The reason why I use fixed point in my resamplers is basically this: I did
many tests (which I can put online for you guys to check) on float vs. fixed
point. The float versions were almost ten times slower, the reason for this
was in around 70% the conversion of the sample pos to integer, for reading
from the sample, and 30% in extracting the decimal part from it to perform
the interpolation.
> With 64bit this would be probably viable because it provides you a
> dynamic range of 385db without scaling and clipping.
> Or is the FPU in the AMD64 just as fast as the integer unit ?
> floating point is easier to deal with (because you do not need to fear
> clipping thus do not need to scale value to emulate
> fixed point fractional arithmetic), but we all know speed is one of the
> main issues in audio software.
>
Regarding to 64 bits, I'd like to add that there is a common misconception
here. All the relevant buses can already transfer 64 bits in nowadays
computers. The FPU and the extensions such as MMX or SSE2 already work
in 64 bits. So there is not really any important gain in speed. The real gain
from this new 64 bits architecture seems to me to be just more registers and
larger adressing...
Juan Linietsky
This archive was generated by hypermail 2b28 : Thu Oct 16 2003 - 22:55:03 EEST