Faster tone

[valdez]

I was looking at the tone detection and noticed if I went to 32 bit math as opposed to the 64 bit, the execution was ~ 1/4 the time.
(I have a small dynamic rage ~ less than 512)

If one reduces bits in the coefficient

eg:
coef = (int32_t)(cos((double)freq
* (2.0 * 3.14159265358979323846 / samples_per_second))
* (double)32767.999);

one would have to also adjust multiplication (and power calculation) as well..

something like :
q0 = (wave_data) + ( ((int32_t)coef * (int32_t) q1)>>14) - q2;


Any suggestions on selecting the correct integer math for 32 bit operation (at the expense of reducing the dynamic range)?

 

[PaulStoffregen]

The large number of bits are needed to detect low frequencies. They're not used for dynamic range of the signal.

 

[valdez]

If one was interested in getting better performance, would there be a way to evaluate a safe minimum frequency (one could down sample if needed, but one would also need to add a high pass filter to be sure one was safe)?