This is somewhat similar to the S32->F32 conversion improvements,
but here things a bit more tricky...
The main consideration is that the limits to which we clamp
must be valid 32-bit signed integers, but not all such integers
are exactly losslessly representable in `float32_t`.
For example it we'd clamp to `2147483647`,
that is actually a `2147483648.0f`,
and `2147483648` is not a valid 32-bit signed integer,
so the post-clamp conversion would basically be UB.
We don't have this problem for negative bound, though.
But as we know, any 25-bit signed integer is losslessly
round-trippable through float32_t, and since multiplying by 2
only changes the float's exponent, we can clamp to `2147483520`!
The algorithm of selection of the pre-clamping scale is unaffected.
This additionally avoids right-shift, and thus is even faster.
As `test_lossless_s32_lossless_subset` shows,
if the integer is in the form of s25+shift,
the maximal absolute error is finally zero.
Without going through `float`->`double`->`int`,
i'm not sure if the `float`->`int` conversion
can be improved further.
There's really no point in doing that s25_32 intermediate step,
to be honest i don't have a clue why the original implementation
did that \_(ツ)_/¯.
Both `S25_SCALE` and `S32_SCALE` are powers of two,
and thus are both exactly representable as floats,
and reprocial of power-of-two is also exactly representable,
so it's not like that rescaling results in precision loss.
This additionally avoids right-shift, and thus is even faster.
As `test_lossless_s32_lossless_subset` shows,
if the integer is in the form of s25+shift,
the maximal absolute error became even lower,
but not zero, because F32->S32 still goes through S25 intermediate.
I think we could theoretically do better,
but then the clamping becomes pretty finicky,
so i don't feel like touching that here.
At the very least, we should go through s25_32 intermediate
instead of s24_32, to avoid needlessly loosing 1 LSB precision bit.
That being said, i suspect it's still not doing the right thing.
Why are we silently dropping those 7 LSB bits?
Is that really the way to do it?
At the very least, we should go through s25_32 intermediate
instead of s24_32, to avoid needlessly loosing 1 LSB precision bit.
FIXME: the noise codepath is not covered with tests.
Tweak the conversion constants a bit so that they handle the
extreme ranges a bit better.
Align the C and vector instructions.
Reactivate the unit test asserts when a conversion fails.
Make a new uint42_t and int24_t type and use that to handle 24 bits
samples. This makes it easier because we can iterate and copy the
structs like other types.
Make dither noise as a value between -0.5 and 0.5 and add this
to the scaled samples.
For this, we first need to do the scaling and then the CLAMP to
the target depth. This optimizes to the same code but allows us
to avoid under and overflows when we add the dither noise.
Add more dithering methods.
Expose a dither.method property on audioconvert. Disable dither when
the target depth > 16.
