On Tuesday, 28 April 2015 at 15:11:16 UTC, Stincky gecko wrote:
> On Tuesday, 28 April 2015 at 15:01:23 UTC, ponce wrote:
>> On Tuesday, 28 April 2015 at 15:00:18 UTC, ponce wrote:
>>> - clearing the state (setting delaylines to 0)
>>> - separate channels that can be stored interleaved or not. Samples would be indexed by time and channel index. I feel like samplerate can be left out of the equation, it will almost always be a runtime value.
>>
>> - and more importantly, batch processing instead of sample per
>> sample like I do now in dplug
>
> in d plug sample by sample processing is justified because of the z-transform, while in video games you just basically play some wavetables.

Even hugely stateful things can benefit of buffering.

For example, biquad filters have an internal loop that looks like (pseudo-code):

    for each sample
    {
        input <- input-signal[n]
        output <- a0 * input + a1 * x[0] + a2 * x[1] - a3 * y[0] - a4 * y[1]

        // update state of filter
        x[1] <- x[0]
        x[0] <- input
        y[1] <- y[0]
        y[0] <- output;
    }

It cannot be parallelized because it's too stateful.
BUT by unrolling it by 3, SIMD can be used and benefit such a loop.

On Tuesday, 28 April 2015 at 15:49:53 UTC, ponce wrote:
> On Tuesday, 28 April 2015 at 15:11:16 UTC, Stincky gecko wrote:
>> On Tuesday, 28 April 2015 at 15:01:23 UTC, ponce wrote:
>>> On Tuesday, 28 April 2015 at 15:00:18 UTC, ponce wrote:
>>>> - clearing the state (setting delaylines to 0)
>>>> - separate channels that can be stored interleaved or not. Samples would be indexed by time and channel index. I feel like samplerate can be left out of the equation, it will almost always be a runtime value.
>>>
>>> - and more importantly, batch processing instead of sample per
>>> sample like I do now in dplug
>>
>> in d plug sample by sample processing is justified because of the z-transform, while in video games you just basically play some wavetables.
>
> Even hugely stateful things can benefit of buffering.
>
> For example, biquad filters have an internal loop that looks like (pseudo-code):
>
>     for each sample
>     {
>         input <- input-signal[n]
>         output <- a0 * input + a1 * x[0] + a2 * x[1] - a3 * y[0] - a4 * y[1]
>
>         // update state of filter
>         x[1] <- x[0]
>         x[0] <- input
>         y[1] <- y[0]
>         y[0] <- output;
>     }
>
> It cannot be parallelized because it's too stateful.
> BUT by unrolling it by 3, SIMD can be used and benefit such a loop.

This is what i meant by z-transform, maybe a confusion in the name, but it think to "the fact that the previous sample value must be stored for processing the next one"...

If you have only a filter it can be batch-processed but if the the filter is plugged to a delay line itself pluged to a SSB frequency shifter whose itself has a feedback bus going to...well it cant be buffered anymore.

In video game, they play wavetables and eventually the output is passed in a binaural spatiializer or a reverb unit...though it's still need sample rate info for re-sampling and maybe dithering depending on the sample bit-depth of the samples frames.

On 29/04/2015 3:27 a.m., Adam D. Ruppe wrote:
> idk if it would be useful at all, I'm not even happy with it myself yet,
> but I am writing a simpleaudio.d
> https://github.com/adamdruppe/arsd/blob/master/simpleaudio.d
>
> it is meant to do waveform i/o and MIDI, but I'm not happy with my first
> draft of the api so it is probably going to change.... eventually.
>
> Currently works on Windows (winMM) and Linux (ALSA).

I would be interested in having a chat with you about what we need long term to get a good audio playing library going.

https://gitter.im/rikkimax

On Tuesday, 28 April 2015 at 17:50:48 UTC, Stincky gecko wrote:
>
> If you have only a filter it can be batch-processed but if the the filter is plugged to a delay line itself pluged to a SSB frequency shifter whose itself has a feedback bus going to...well it cant be buffered anymore.
>

I see what you mean, but maybe it can be buffered.

With ae.utils.graphics, even a long chain of operations still benefits from pixel being arranged in scanlines. It's like "expression templates for images" and maybe we need "expression templates for sound".

The memory support for this is often alloca() in the C++ world. And it's pretty cumbersome to do everything with buffers.

> In video game, they play wavetables and eventually the output is passed in a binaural spatiializer or a reverb unit...though it's still need sample rate info for re-sampling and maybe dithering depending on the sample bit-depth of the samples frames.

Indeed wavetable look-up is not stateful.