Phase1b of audio engine rewrite

[yara-blue]

review after #903 (this includes it's changes)

Tracking issue: #901

• Expands the fixed source impl. to match const source
• de-duplicates methods between them using a new common::source module and macros
• makes a note in the changelog
• uses Placeholder to have docs and examples point to not yet landed code

[roderickvd]

I see this is built on top of #903 so I didn't duplicate any of my review points there to here.

[roderickvd]

This one and IntoDynamicSource should forward try_seek.

[roderickvd]

Should forward try_seek.

[roderickvd]

Consider exporting this one so users can downcast to it.

[roderickvd]

Should re-add size_hint().

[roderickvd]

This seems misplaced or having the wrong name.

[yara-blue]

how so? it is the documentation for FixedSource::collect_into_buffer and ConstSource::collect_into_buffer.

I'll expand the text a little to make that more clear.

[roderickvd]

I thought it was out of place because the title was about building a SamplesBuffer, not about collecting into a buffer, and went on to describe a panic that wouldn't be occurring in new (the "builder") anymore, but in total_duration.

[roderickvd]

As per #913 we should add the ability to create one with an arbitrary channel count. Over there we targeted it for phase 4, so just mentioning it here.

[yara-blue]

For silence the idea is to always put it through a channel convertor. My hope being that the optimizer can figure it out. If that is false and perf. becomes an issue we could specialize it.

[roderickvd]

Automatic optimization would be great, but how would that work?

[yara-blue]

the compiler knows that Silence::next returns a constant value (namely: Some(0.0)).

Lets take this fragment of the channel convertor:

        let result = match self.next_output_sample_pos {
            0 => {
                let value = self.input.next();
                self.sample_repeat = value;
                value
            }
            x if x < self.from.get() => self.input.next(),
            1 => self.sample_repeat,
            _ => Some(Sample::EQUILIBRIUM),
        };

Every self.input.next() gets replaced with Some(0):

        let result = match self.next_output_sample_pos {
            0 => {
                self.sample_repeat = 0.0;
                0.0
            }
            x if x < self.from.get() => 0.0,
            1 => self.sample_repeat,
            _ => Some(0.0),
        };

Now its simple to see that self.sample_repeat is also constant. So everywhere we see it replace with a 0.0.

        let result = match self.next_output_sample_pos {
            0 => {
                0.0
            }
            x if x < self.from.get() => 0.0,
            1 => 0.0,
            _ => Some(0.0),
        };

And this is now just a match statement returning 0.0 always so result can be set to 0.0.

There is some other stuff like self.from.get() and self.next_output_sample_pos() but that has also become dead code so the optimizer throws that out as well. And now we've gotten to the point where ChannelCountConvertor::next() can be replaced with just 0.0. This should domino up any audio pipeline.

Now if you put silence in a box and cast it to a Box... then this stops working.

Whether this all happens? It probably does, these are basic optimizations. You can never know for sure, sometimes a suggestion here and there is needed (a little #[inline] usually).

[roderickvd]

Ah, thanks for that write-up, that makes sense. As long as it doesn't get fed into Rubato or something, because then all bets would be off I imagine.