This allows us to change the volume of sounds in filters ("volume",
"fade" properties) without quantizing the audio prematurely. Previously,
we had a hack to apply volume changes as a mixing-time step. This
can be done at any point in the filter chain now.
This also means we only need to worry about clipping at the very
last mixing step, instead of at every place we change sound.
Dithering for 16-bit audio and more intelligent gain control
is possible, as we have full-resolution, unclipped sound at the
mixing phase.
We support high resolution source sounds; both Vorbis and MAD
supply more than 16 bits.
Several filters, such as resampling, are easier to implement as
floats; these no longer need to convert back and forth.
Negatives:
- More memory use. The main case of this is in RageSoundReader_Preload,
which will be fixed: we can preload in 16-bit without losing most of the
above.
- Some extra overhead for accessing more memory.
aligning their speeds very difficult. The resampler is more limited, and it's easier
to remove this restriction from the speed changer, so reverse this logic, basing
the ratio on the resampler instead of the speed changer. (also simpler)
Glenn Maynard