[fluid-dev] Floats and doubles, simd and interpolation

[David Henningsson]

Just a short dump of what's been on my mind the last few weeks...

Yesterday, I tried to chase down a bug I discovered - I rendered a song 
but a few notes were slightly out of tune, but only when using floats, 
not when using doubles! I never bought that doubles should sound better 
than floats - floats has 24 bits of mantissa precision, that's the best 
you can get out of a soundcard anyway, and since all 24 bits soundcards 
have a noise floor that make them no more than 20 bits anyway (right?), 
floats should be enough, even given some rounding errors. (Not to 
mention all the analog components after the DAC...)
I also doubt that anyone would hear the difference between 440.000 Hz 
and 440.001 Hz, if the error is that way. So it must be a bug somewhere.

While looking at that I saw that the interpolation routines has 
conversions to and from double (not fluid_real_t), so I suspect that 
could be why some people claim that doubles are faster than floats in 
some cases.

So the reason I like floats is that with SSE, you can process 4 floats 
simultaneously, but only 2 doubles. From running a perf I know that 2/3 
of the time (for my testcase) was spent in the interpolation routine. 
Can we SIMD:ize that, we might get 3-4x speed improvement, that's at 
least what I hope for.

There is a library called "ORC", anybody heard of it? You write some 
pseudo-assembly code, and on first run ORC translates it into SSE, MMX, 
Altivec, etc, or plain old C depending on your hardware. I think it 
sounds interesting, and was hoping to see if I could make a test soon, 
but then I got busy trying to find that bug instead.

// David