Each 4 related all-passes now share a structure with one delay line, which uses
an interleaved sample history. Also fixes some potential rounding problems for
delay lines with interleaved samples.
This seems to be more in-line with the intended behavior, to allow build-up and
overlap within the reverb decay, rather than a pitch-shift on input.
Unfortunately there's no readily available implementation of this reverb model
that includes modulation to compare with, so a low depth coefficient is used to
keep it very subtle.
The previous value couldn't actually be expressed as a float and got rounded up
to the next whole number value, leaving the potential for an overrun in the
squared sum.
This is a bit more efficient than calling the normal HRTF mixing function
twice, and helps solve the problem of the values generated from convolution not
being consistent with the new HRIR.
Some data sets are just too sparse, having noticeably few measurements to
properly handle slowly panning sources. Although not perfect, bilinearly
interpolating the HRIR measurements improves the positional accuracy.
This greatly improves HRTF performance since the dual-mix only applies to the
64-sample coefficient transition. So rather than doubling the full mix, it only
doubles 64 samples out of the full mix.
This reduces the output volume when the mixed samples extend outside of -1,+1,
to prevent excessive clipping. It can reduce the volume by -80dB in 50ms, and
increase it by +80dB in 1s (it will not go below -80dB or above 0dB).
This update modifies the reverb in numerous ways. The 3-series, 4-parallel
all-pass is replaced with a Gerzon vector all-pass. The vector all-pass is also
applied to the early reflections, to help with the initial diffusion in the
late reverb. The late reverb filter and feedback attenuation has been replaced
with a dual first-order equalization filter, which controls the low and high
frequencies with individual low-pass/high-shelf and high-pass/low-shelf filters
with gain control.
Additionally, delay lines now have the ability to blend the output between two
offsets, to help smooth out the transition for when the delay changes (without
such, it could result in undesirable clicks and pops).
Clang does not allow using C11's atomic_load on const _Atomic variables.
Previously it just disabled use of C11 atomics if atomic_load didn't work on a
const _Atomic variable, but I think I'd prefer to have Clang use C11 atomics
for the added features (more explicit memory ordering) even if it means a few
instances of breaking const.
