[openal] HRTF redesign

[Chris Robinson]

Hey guys.

I've been giving some serious thought to redesigning how HRTF mixing 
works. Currently, it works in a fairly straight-forward, if somewhat 
costly, way... when a source updates, its direction is used to lookup 
the filter coefficients (really, the four nearest sets for bilinear 
interpolation) which are then used to apply a FIR filter for the source 
input to two output channels. There's also per-sample stepping to fade 
between coefficients as the source pans around and the attenuation changes.

A possible alternate way, I think, is by mixing normally to a virtual 
14-channel setup (8-channel cube + 6-channel "diamond"), then apply HRTF 
to those channels for output. This has the obvious drawback that it 
loses some positional accuracy, but it does have a few benefits.

First, it simplifies HRTF setup since it only needs to calculate 
coefficients once on device reset, and avoids any need for per-sample 
stepping. Secondly, HRTF processing then becomes a fixed cost. The HRTF 
filter applies to the 14 channels, irrespective of however many sources 
are playing. It also affects the output of effects.

As well, it improves B-Format handling since it would mix into 14 
distinct channels around the listener, which then get filtered for a 
more encompassing result instead of simple stereo.

But... I'm not sure how bad the positional accuracy loss will be, 
especially given the ambisonics-based panning which would spread the 
signal out across multiple (virtual) channels. Though maybe the way the 
channel volumes are balanced will still provide good positioning cues. 
So I'm curious if anyone has any thoughts on this.