92 lines
3.5 KiB
Python
92 lines
3.5 KiB
Python
|
# Copyright 2024-2025 XMOS LIMITED.
|
||
|
|
# This Software is subject to the terms of the XMOS Public Licence: Version 1.
|
||
|
|
import numpy as np
|
||
|
|
import pytest
|
||
|
|
from pathlib import Path
|
||
|
|
|
||
|
|
import audio_dsp.dsp.fir as fir
|
||
|
|
import audio_dsp.dsp.signal_gen as sg
|
||
|
|
import audio_dsp.dsp.utils as utils
|
||
|
|
from test.test_utils import q_convert_flt, assert_allclose
|
||
|
|
|
||
|
|
gen_dir = Path(__file__).parent / "autogen"
|
||
|
|
|
||
|
|
# Note the filter coeffs files are defined in test/fir/conftest.py
|
||
|
|
@pytest.mark.parametrize("coeff_path", ["passthrough_filter.txt",
|
||
|
|
"descending_coeffs.txt",
|
||
|
|
"simple_low_pass.txt",
|
||
|
|
"aggressive_high_pass.txt",
|
||
|
|
"comb.txt",
|
||
|
|
"tilt.txt"])
|
||
|
|
def test_basic(coeff_path):
|
||
|
|
fut = fir.fir_direct(48000, 1, Path(gen_dir, coeff_path))
|
||
|
|
|
||
|
|
signal = sg.pink_noise(48000, 0.1, 0.5)
|
||
|
|
signal = q_convert_flt(signal, 23, 27)
|
||
|
|
# signal = np.zeros(1000)
|
||
|
|
# signal[0] = 1
|
||
|
|
|
||
|
|
coeffs = np.loadtxt(Path(gen_dir, coeff_path))
|
||
|
|
out_ref = np.convolve(signal, coeffs)[:len(signal)]
|
||
|
|
|
||
|
|
out_flt = np.zeros_like(signal)
|
||
|
|
out_int = np.zeros_like(out_flt)
|
||
|
|
|
||
|
|
for n in range(len(signal)):
|
||
|
|
out_flt[n] = fut.process(signal[n])
|
||
|
|
|
||
|
|
fut.reset_state()
|
||
|
|
|
||
|
|
for n in range(len(signal)):
|
||
|
|
out_int[n] = fut.process_xcore(signal[n])
|
||
|
|
|
||
|
|
# difference in convolution implementations means flt and ref aren't
|
||
|
|
# bit exact, especially after saturation!
|
||
|
|
unsaturated = ((out_ref > float(-(2 ** (31 - fut.Q_sig)))) &
|
||
|
|
(out_ref < float((2**31 - 1) / 2**fut.Q_sig)))
|
||
|
|
np.testing.assert_allclose(out_flt[unsaturated], out_ref[unsaturated], atol=2**-52)
|
||
|
|
|
||
|
|
# small signals are always going to be ropey due to quantizing, so just check average error of top half
|
||
|
|
top_half = utils.db(out_flt) > -100
|
||
|
|
if np.any(top_half):
|
||
|
|
error_flt = np.abs(utils.db(out_int[top_half])-utils.db(out_flt[top_half]))
|
||
|
|
mean_error_int = utils.db(np.nanmean(utils.db2gain(error_flt)))
|
||
|
|
assert mean_error_int < 0.016
|
||
|
|
np.testing.assert_allclose(out_flt, out_int, atol=2**(-21))
|
||
|
|
|
||
|
|
|
||
|
|
# Note the filter coeffs files are defined in test/fir/conftest.py
|
||
|
|
@pytest.mark.parametrize("coeff_path", ["passthrough_filter.txt",
|
||
|
|
"descending_coeffs.txt",
|
||
|
|
"simple_low_pass.txt"])
|
||
|
|
@pytest.mark.parametrize("n_chans", [1, 2, 4])
|
||
|
|
def test_frames(coeff_path, n_chans):
|
||
|
|
fut = fir.fir_direct(48000, n_chans, Path(gen_dir, coeff_path))
|
||
|
|
|
||
|
|
signal = sg.pink_noise(48000, 0.1, 0.5)
|
||
|
|
signal = np.tile(signal, [n_chans, 1])
|
||
|
|
signal[0] = -signal[0]
|
||
|
|
frame_size = 1
|
||
|
|
|
||
|
|
signal_frames = utils.frame_signal(signal, frame_size, 1)
|
||
|
|
|
||
|
|
out_flt = np.zeros_like(signal)
|
||
|
|
out_int = np.zeros_like(out_flt)
|
||
|
|
|
||
|
|
for n in range(len(signal_frames)):
|
||
|
|
out_flt[:, n*frame_size:(n+1)*frame_size] = fut.process_frame(signal_frames[n])
|
||
|
|
assert np.all(-out_flt[0, :] == out_flt[1:, :])
|
||
|
|
fut.reset_state()
|
||
|
|
|
||
|
|
for n in range(len(signal_frames)):
|
||
|
|
out_int[:, n*frame_size:(n+1)*frame_size] = fut.process_frame_xcore(signal_frames[n])
|
||
|
|
|
||
|
|
for n in range(1, n_chans):
|
||
|
|
# rounding differences can occur between positive and negative signal
|
||
|
|
assert_allclose(-out_int[0, :], out_int[n, :], atol=(2**(-fut.Q_sig)))
|
||
|
|
|
||
|
|
|
||
|
|
if __name__ =="__main__":
|
||
|
|
# test_basic("simple_low_pass.txt")
|
||
|
|
# for n in range(100):
|
||
|
|
test_frames("simple_low_pass.txt", 2)
|