DSP optimizations

[danomatika]

This external performs DSP convolution by basically doing a lot of matrix math on a 368x2x128 data set and can get a bit CPU heavy when using multiple objects.

Some optimizations could be:

• basic: change the main for loops to use pointer incrementation as opposed to indexing
• advanced: platform/arch-specfic optimized matrix & vector math libs (for Apple platforms, the Accelerate framework)

ping @SylvainPDR

[chikashimiyama]

vDSP for mac, MTL for win, KissFFT for linux? platform/arch-specific approach costs a lot of dev time.
Platform independent options:
fftw http://fftw.org/
kissFFT could be the easiest though.

[danomatika]

Whatever is cross-platform and easiest. Pierre (intern) looked into this and said there wasn't much he could change to make it faster, but it's worth a second look just in case.

[chikashimiyama]

ok.

[chikashimiyama]

earplug/earplug~.c

Lines 151 to 178 in 4ac1938

	while (blocksize--)
	{
	convSum[0] = 0;
	convSum[1] = 0;
	inSample = *(in++);
	x->convBuffer[x->bufferPin] = inSample;
	unsigned scaledBlocksize = blocksize * blockScale;
	unsigned blocksizeDelta = 8191 - scaledBlocksize;
	for (i = 0; i < 128; i++)
	{
	convSum[0] += (x->previousImpulse[0][i] * x->crossCoef[blocksizeDelta] +
	x->currentImpulse[0][i] * x->crossCoef[scaledBlocksize]) *
	x->convBuffer[(x->bufferPin - i) &127];
	convSum[1] += (x->previousImpulse[1][i] * x->crossCoef[blocksizeDelta] +
	x->currentImpulse[1][i] * x->crossCoef[scaledBlocksize]) *
	x->convBuffer[(x->bufferPin - i) &127];
	x->previousImpulse[0][i] = x->currentImpulse[0][i];
	x->previousImpulse[1][i] = x->currentImpulse[1][i];
	}
	x->bufferPin = (x->bufferPin + 1) & 127;
	*left_out++ = convSum[0];
	*right_out++ = convSum[1];
	}
	return w + 6;

convolution code

[chikashimiyama]

L. 170 - 171 is totally redundant code. and possibly this is not intended. since this is done in the nested loop of while(block size--) and for. the intended crossfading may be not working at all.

[chikashimiyama]

@danomatika

I tried frequency domain convolution using uFFT but the result is not very good (not so optimized significantly and generates some artifacts). see the optimization branch.
I can investigate more but I'm not sure if I should use more time for this...

[danomatika]

This refers to PR #16