Hi again Dan,

Following our previous discussion on splitting audio, I’m working on a means of managing LFE level after redirecting the bass from the 7 main channels. Rather than just attenuate the LFE samples by a fixed number of dB, I’m implementing dynamic range compression. I use a Mac Mini 2018 and my KFR processing is all double precision floating point. I have plenty of amplifier power and subs to deliver high sound pressure levels at low frequencies. :)

The compression algorithm I’m using (for better or worse), is based on a circular arc. For a sample s, the function drc(s) is:
if abs(s) <= threshold (say 0.9), drc(s) = s.
if threshold < abs(s) < upper limit, drc(s) = point on a circular arc centred such that drc() is differentiable at all values of s. Upper limit is set at the point where the arc has zero slope. Soft clip.
if abs(s) > upper limit, drc(s) = 1.0. Hard clip.
Sorry if this is a bit cryptic.

The function’s code looks like this (NB calculation of some constants is not shown):
double grc(double apply_circular_arc_drc(double s)
if (abs(s) <= threshold)
return s; // Sample is below threshold; no compression.
else if (abs(s) < s_input_max) {
// Sample is above threshold and below upper limit
const double w = s_input_max - abs(s); // Find width of triangle for sample
// Calculate height of point on circular arc on which the sample lies using Pythagoras
const double h = std::pow((std::pow(r, 2) - std::pow(w, 2)), 0.5);
// Calculate sample as height of triangle in circle plus height of centre of circle above 0.0
const double t = (h + c) * sign(s); // Retain sign of original sample
return t;
} else {
// Sample is above upper limit; hard clip.
double t = 1.0 * sign(s); // Retain sign of original sample
return t;
}

I apply the compression to univector samples using:
std::transform(samples.begin(), samples.end(), samples.begin(), g);

This runs the function against every sample in univector and does not leverage simd intrinsics. It runs fine, but I'm looking for a more efficient solution.
Question: How could this be implemented more efficiently using KFR?
I’ve tried to work out whether I can use KFR template expressions (and read the example) and the existing sqr and sqrt functions but haven’t found a solution.

Appreciate your help with this.

PS Pls let me know if this approach is nonsense and should be done another way!