I came across this ESP FFT library : FFT for onboard ESP32 It performs quite well, few miliisecs for a 512-long signal
It is written as a plain C function. To prevent repeated malloc and free's for each FFT run, I decided to convert it into a c++ class. It also improves code readability I added a Hamming window function, because windowing is pretty much always needed for any FFT.
The above link provides all docu for FFT in general, the repo says it all
The way it is used:
#define FFT_N 2048 // The length of the FFT. Must be a power of 2
#define SAMPLEFREQ 8192 //
float fft_output[FFT_N];
#include "ESP_fft.h"
#include "fft_signal.h"
// Create the class instance
// Typical embedded FFT use is forward (sigal->frequencies) and real (get each freq & amplitude)
// use the signal from our file
//
ESP_fft FFT(FFT_N, SAMPLEFREQ, FFT_REAL, FFT_FORWARD, samples, spectrum);
void setup() {
Serial.begin(115200); // use the serial port
long int t1 = micros();
// Execute transformation
// Window the signal. See theory: without windowing, the tranition from nothing to signal at the beginning,
// and signal to nothing at the end will cause the presence of frequencies that we don;t actually have
//
FFT.hammingWindow();
// optional:
// after Hamming, usually some DC (offset from average) is introduced in the signal
// with removeDC (which is in fact 'make AC'), average is subtracted so that signal centers around 0
// you can safely skip this step, the FFT 'formula' puts DC value in output[0]
FFT.removeDC();
// then do the actual FFT
FFT.execute();
// we usually want the amplitudes of each frequency, so tis step is required
//
FFT.complexToMagnitude();
long int t2 = micros();
Serial.print("Time taken: ");Serial.print((t2-t1)*1.0/1000);Serial.println(" milliseconds!");
Serial.println();
// Below is what the original doc states, but I found by testing
// (using parseval's law, comparing power/RMS of signal and spectrum after removing DC) that
// this is quite a different value. Anyway, needs quite a bit more work
// and in practice the actual amplitude / power in the frequency domain is not really that important.
/*Multiply the magnitude of the DC component with (1/FFT_N) to obtain the actual DC component*/
Serial.printf("DC component : %f g\n", (spectrum[0])/FFT_N); // DC is at [0]
/*Multiply the magnitude at all other frequencies with (2/FFT_N) to obtain the amplitude at that frequency*/
Serial.printf("Fundamental Freq : %f Hz\t Mag: %f g\n", FFT.majorPeakFreq(), (FFT.majorPeak()/10000)*2/FFT_N);
for (int i=0; i< 10; i++) {
Serial.printf("%f:%f\n", FFT.frequency(i),spectrum[i]);
}