mcid.py: optimize FFT and A-weighting calculations (#33057)

* Precomputing weighting * add comments back * use cache * spacing spacing * clean up * lower by diff --------- Co-authored-by: Shane Smiskol <[email protected]>
commaai · Jul 25, 2024 · 313a282 · 313a282
1 parent 2da4aef
commit 313a282
Show file tree

Hide file tree

Showing 2 changed files with 13 additions and 11 deletions.
diff --git a/selfdrive/test/test_onroad.py b/selfdrive/test/test_onroad.py
@@ -56,7 +56,7 @@
  "system.logmessaged": 0.2,
  "system.tombstoned": 0,
  "logcatd": 0,
- "system.micd": 6.0,
+ "system.micd": 5.0,
  "system.timed": 0,
  "selfdrive.pandad.pandad": 0,
  "system.statsd": 0.4,

diff --git a/system/micd.py b/system/micd.py
@@ -1,5 +1,6 @@
 #!/usr/bin/env python3
 import numpy as np
+from functools import cache
 
 from cereal import messaging
 from openpilot.common.realtime import Ratekeeper
@@ -10,7 +11,16 @@
 FFT_SAMPLES = 4096
 REFERENCE_SPL = 2e-5 # newtons/m^2
 SAMPLE_RATE = 44100
-SAMPLE_BUFFER = 4096 # (approx 100ms)
+SAMPLE_BUFFER = 4096 # approx 100ms
+
+
+@cache
+def get_a_weighting_filter():
+ # Calculate the A-weighting filter
+ # https://en.wikipedia.org/wiki/A-weighting
+ freqs = np.fft.fftfreq(FFT_SAMPLES, d=1 / SAMPLE_RATE)
+ A = 12194 ** 2 * freqs ** 4 / ((freqs ** 2 + 20.6 ** 2) * (freqs ** 2 + 12194 ** 2) * np.sqrt((freqs ** 2 + 107.7 ** 2) * (freqs ** 2 + 737.9 ** 2)))
+ return A / np.max(A)
 
 
 def calculate_spl(measurements):
@@ -27,16 +37,8 @@ def apply_a_weighting(measurements: np.ndarray) -> np.ndarray:
  # Generate a Hanning window of the same length as the audio measurements
  measurements_windowed = measurements * np.hanning(len(measurements))
 
- # Calculate the frequency axis for the signal
- freqs = np.fft.fftfreq(measurements_windowed.size, d=1 / SAMPLE_RATE)
-
- # Calculate the A-weighting filter
- # https://en.wikipedia.org/wiki/A-weighting
- A = 12194 ** 2 * freqs ** 4 / ((freqs ** 2 + 20.6 ** 2) * (freqs ** 2 + 12194 ** 2) * np.sqrt((freqs ** 2 + 107.7 ** 2) * (freqs ** 2 + 737.9 ** 2)))
- A /= np.max(A) # Normalize the filter
-
  # Apply the A-weighting filter to the signal
- return np.abs(np.fft.ifft(np.fft.fft(measurements_windowed) * A))
+ return np.abs(np.fft.ifft(np.fft.fft(measurements_windowed) * get_a_weighting_filter()))
 
 
 class Mic: