Block-Based Processing in Third Party Node

[A Former User]

Sorry for the dumb question, had a bit of brain fog the last few days

template <typename T> void process(T& data)
		{
			static constexpr int NumChannels = getFixChannelAmount();
			// Cast the dynamic channel data to a fixed channel amount
			auto& fixData = data.template as<ProcessData<NumChannels>>();

			// Create a FrameProcessor object
			auto fd = fixData.toFrameData();

			while (fd.next())
			{
				// Forward to frame processing
				processFrame(fd.toSpan());
			}

		}

Where in here do we process full buffers of audio? I know we can do single sample processing but I'm not sure how to do block-based processing

[Christoph Hart]

// ch is a pointer to the begin of each channel
for(auto ch: data)
{
	// This gets you the array to the channel signal
	dyn<float> channel = data.toChannelData(ch);

	// access any sample
	channel[2] = 0.0f;

	// perform math operation on the entire
	// data (using SIMD where possible)
	channel *= 2.0f;

	// iterate over the sample
	for(auto& sample: channel)
		sample *= 0.5;
}

[A Former User]

@Christoph-Hart thank you, I'll give it a shot.

sidenote I almost have RTNeural working in HISE, if I can get this bit up and running I'll make a post in case anyone else wants to try out the ol neural synthesis :)

[A Former User]

Hmm okay block-based & single sample processing achieve the same result, is there a method to push samples to a dyn array in a FIFO manner? Would that even be stable in the audio thread?

[Christoph Hart]

Nice. If you need to pass in the channel pointers to an external function from a library, there is also

void someExternalStuff(float** ptrs, int numSamples, int numChannels)
{

}


float** ptrs = data.getRawDataPointers();
int numSamples = data.getNumSamples();
int numChannels = data.getNumChannels();

[A Former User]

Sweet. RTNeural also supports SIMD, I assume that would play nicely with HISE?

Have you come across any compatibility issues with SIMD?

[Christoph Hart]

SIMD is just a way of calculating numbers and HISE (or any other host application) doesnt‘t care how a library produces its output.

Usually you just get the pointers and sample amount and pass that to the processing function of your library.

[A Former User]

@Christoph-Hart said in Block-Based Processing in Third Party Node:

dyn<float> channel = data.toChannelData(ch);
float** ptrs = data.getRawDataPointers();

Is there a way to cast or have these return as doubles?

[Christoph Hart]

No.

Can you set the library to use float numbers as signal? Usually every library allows to define the floating point type used for its signal (eg. Faust is using FAUST_FLOT which you can define to either float (reasonable) or double ("yeah, 64bit signal precision for the ultimate transparent sound").

If you need to feed it double numbers then it will become very annoying because you can't just cast the pointers to double pointers because the original source are still float numbers and you would have to introduce a working buffer that uses double numbers, copy the float numbers over, process the working buffer with your library and reverse all those steps to copy it back to the signal buffer.

[A Former User]

Yeah RTNeural supports floats. I think the issue is with the type of Model I'm using -> I can get a guitar amp black box model working just fine, but the whole "predict the future" version isn't playing very nice :)

I'll keep messing around