Fix AudioStream cubic resampling

[ellenhp]

Reverts b2264cb and fixes Godot's cubic resampling algorithm. Lots of context for this is in #23544 and #49131.

Fixes #49131.

This is implemented with some help from the math on this page: https://www.paulinternet.nl/?page=bicubic

I'm not really a math person. :)

[fire]

Reviewing with Ellen offline.

[Zylann]

Comments here #23544 (comment)

[ellenhp]

Summarizing discussion, I think the issue Zylann is running into is related to the WASAPI driver, not AudioStreamPlaybackResampled bugs. I'll probably try and come up with a testing matrix and use OBS to record what everything sounds like so people can review without trying every possible test case.

[ellenhp]

Okay, I have a demo of this now! Took a while fiddling with OBS but I did it.

The audio files I used are in this zip:

soundfiles.zip

My system's sample rate is 44100, so keep that in mind here. The other thing is that the github video player seems to start muted so turn on the audio for these :)

---

This is Godot 3.2.3 with the project mix rate set to 44100. Everything sounds fine, nothing is unusual about it.

godot3_official_44100.mp4

This is Godot 3.2.3 with the project mix rate set to 48000. There's a fairly audible ringing artifact on the low frequency beat sound. This is what I fixed with #46086

godot3_official_48000.mp4

Fast-forward to today, with a build from 23bf04a (sample rate 44100). Theoretically there's an audible low-pass effect (#49131) from the new resampling algorithm, but I can't hear it with these audio files.

godot4_head_44100.mp4

And 48000. Should also have a low-pass.

godot4_head_48000.mp4

So what I did in this PR is reverted to the old cubic algorithm and adjusted the coefficients to match a derivation I found online (linked above). Here's how that sounds at 44100 (theoretically a no-op).

godot4_fixcubic_44100.mp4

And here's how it sounds at 48000, which caused issues prior to the coefficient adjustment.

godot4_fixcubic_48000.mp4

No ringing! At least that I can hear.

[akien-mga]

I've tested the changes locally with the files provided in #51082 (comment), and my bad ears and cheap earplugs didn't let me notice much difference (on Linux with Pulseaudio, system sampling rate seems to be 48000 Hz). No regression either though, and since cubic resampling seems to be the better option, it definitely makes sense to go back to it with a fixed implementation.

I can confirm that the new implementation does not reintroduce the ringing from #23544.

I suggest we merge this, include it in the 3.4 betas, and we'll see if audiophiles find new issues or can confirm that the old issues have been fixed.

[akien-mga]

Thanks!

[akien-mga]

Cherry-picked for 3.4. (Will be included in upcoming 3.4 beta 3.)

[akien-mga]

Cherry-picked for 3.3.3.