Making autocorrelation more efficient about pitch-detection HOT 3 CLOSED

I'll explore this with some benchmarks.

from pitch-detection.

for (int i = 0; i < N2/2; ++i) {
out[i] = outa[i] * std::conj(outb[i]) * scale;
out[N2-1 - i ] = std::conj(out[i]);
}

The second line there, out[N2-1 - i] = std::conj(out[i]); is causing the results to be junk. However, you are onto something. Since at the bottom of MPM, I only take the first N results of the FFT:

	std::vector<double> normalized_result(N, 0.0);
	for (int i = 0; i < N; ++i)
		normalized_result[i] =
		    std::real(result[i + (N2 - N)]) / std::real(result[N2 - N]);

I follow your suggestion to cut the conj * scale loop in half:

	std::complex<float> scale = {1.0f / (float)N2, 0.0};
	for (int i = 0; i < N; ++i) 
		out[i] = outa[i] * std::conj(outb[i]) * scale;

This gives the identical detected pitch outputs with better performance (quite obvious the reason why).

from pitch-detection.

477b095

from pitch-detection.