0
$\begingroup$

The Broadcom BCM2711 GPIO controller used by the Raspberry Pi has a hardware PWM implementation option using an algorithm that provides an average value defined by a rational fraction $N/M$ defined as follows in section 8.3. PWM Implementation starting on page 156:

  1. Set context = 0
  2. context = context + N
  3. if (context >= M) context = context - M send 1 else send 0
  4. Repeat from step 2

where context is a register which stores the result of the additions/subtractions.

I've implemented this in a Python script to try to start to get an idea how I'll need to filter it in order to get a low-noise, stable voltage. Here with $M=200$ I've incremented through $0 \le N \le M$ for over 1000 time points each, then applied a gaussian smoothing with a $\sigma$ of 10 time points just to get a quick look.

I took the Fourier transform of the middle 1024 points, ignoring a $3 \sigma$ buffer on each side.

I was surprised to see that some values of $N$ produce signals that are orders of magnitude more noisy than other nearby values of $N$.

Questions:

  1. This kind of analysis is new territory for me, if I've made any major errors in the analysis I would appreciate knowing about it.
  2. If I need near-DC voltages I can just run the BCM2711 at a moderately high frequency and filter the heck out of the output with a simple RC filter. But if I wanted to work closer to the frequency that the algorithm was running, what additional analysis should I do on this algorithm before deciding on a more aggressive low-pass filter and simulating it's performance?

filtered waveforms for different values of N

power spectra of filtered waveforms

stdev of filtered waveforms

Here's the Python script for the plots and analysis:

import numpy as np
import matplotlib.pyplot as plt
from scipy import ndimage
from numpy.fft import rfft, rfftfreq

sigma = 10
ignore = 3 * sigma
M = 200
big = []
for N in range(M+1):
    context = 0
    bob = []
    for i in range(1024 + 2 * ignore):
        context += N
        if (context >= M):
            context -= M
            bob.append(1)
        else:
            bob.append(0)
    big.append(bob)

big = np.array(big, dtype=float)
fbig = ndimage.gaussian_filter1d(big, sigma=sigma, mode='mirror', order=0, axis=-1)
fbig_trunc = fbig[:, ignore:-ignore]
print('fbig_trunc.shape: ', fbig_trunc.shape)
ft = rfft(fbig_trunc, axis=-1)
power = np.abs(ft)**2
power /= power.max()
dB_rel = 10 * np.log10(power)
print('power.shape: ', power.shape)
print('power.max(): ', power.max())
print('dB_rel.shape: ', dB_rel.shape)
print('dB_rel.max(): ', dB_rel.max())

if True:
    plt.figure()
    for thing in fbig:
        plt.plot(thing, linewidth=0.8)
    plt.ylim(-0.02, 1.02)
    plt.title('sigma={:2d}, M={:3d}'.format(sigma, M), fontsize=16)
    plt.xlabel('time steps', fontsize=16)
    plt.show()

if True:
    plt.figure()
    maxi = dB_rel[:, 1:201].max()
    plt.imshow(dB_rel[:, :201], vmax=maxi, vmin=-100)
    plt.colorbar()
    plt.gca().set_aspect(2.0)
    plt.title('FT relative power (dB), sigma={:2d}, M={:3d}'.format(sigma, M), fontsize=16)
    plt.xlabel('frequency', fontsize=16)
    plt.ylabel('N', fontsize=16)
    plt.show()

if True:
    plt.figure()
    std = fbig_trunc.std(axis=-1)
    plt.plot(10*np.log10(std**2))
    plt.title('stdev relative power, sigma={:2d}, M={:3d}'.format(sigma, M), fontsize=16)
    plt.xlabel('N', fontsize=16)
    plt.ylabel('stdev (dB rel)', fontsize=16)
    plt.show()
$\endgroup$
1
$\begingroup$

The pulse generation algorithm sounds like it is more like Pulse Density Modulation (PDM), not Pulse Width Modulation (PWM), because PWM would pack the ones into the beginning of counting, and leave the zeroes in the end.

Indeed some values of N and M cause the output pattern to have a certain repetitive bit stream, for example the 100/200 means the output is 50% 0 and 50% 1, so the output is a bitstream of 1,0,1,0... which is easy to filter. If you set N and M to be exactly ratio of 1/3, the output pattern will also be repetitive, 0,0,1,0,0,1... So, in light of this this also explains the spectrogram of N=50 M=200, which is 25%, 0,0,0,1,0,0,0,1.

These patterns that have a short repetition rate show up as pure tone peaks in the spectrogram. Choosing the values so that there is no obvious short repeating patterns basically dithers the bitstream and widens the spectrum to have multiple lower peaks.

Depending on what is the usage of the output will determine how to filter it. For many cases a simple RC filter is enough, as long as the PDM bitstream output rate is high enough for the RC filter cutoff frequency.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Ah! This is extremely helpful; searching for Pulse Density Modulation opens up a whole body of literature for me to read, thank you! $\endgroup$ – uhoh May 17 at 9:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.