Derive sin/cos of PLL phase error directly from I/Q arms (q/mag, i/mag)
instead of calling atan2 + sin_cos. Use double-angle identity to compute
38 kHz carrier (sin2θ = 2·sinθ·cosθ, cos2θ = 2·cos²θ-1) from the
rotated pilot sin/cos, eliminating the second sin_cos call entirely.
Drop Butterworth from 6th to 4th order (resampler Blackman-Harris now
handles stopband). Use power-of-2 bitmask for ring buffer indexing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Halves the coefficient bank from 128×32 to 64×32 (16 KB → 8 KB) for
better L1 cache utilization while maintaining sufficient fractional
sample resolution.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Replace shift_append (O(N) rotate_left per sample) with a circular buffer
index for O(1) writes. The polyphase resampler now reads from the ring
buffer directly, eliminating 3 × 32-element memmoves per composite sample.
Remove unused dot_product functions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Pre-compute all FM discriminator outputs using demod_fm_with_prev which
processes 8 samples at a time via AVX2 atan2, then iterate the scalar
results through the rest of the stereo pipeline. Eliminates per-sample
f32::atan2 calls from the inner loop.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
With Blackman-Harris window and proper cutoff (~0.24), 32 taps still
provides 60+ dB stopband rejection. Halves the per-sample MAC count
from 192 to 96 across the three resampler channels.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Increase polyphase resampler phases from 32 to 128 for finer fractional
sample positioning. Replace Hamming window with Blackman-Harris for ~92 dB
stopband rejection. Add pilot notch on composite signal before diff demod
to prevent 19 kHz × 38 kHz intermod products in the stereo difference
path.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Increase polyphase resampler taps from 16 to 64 for sharper anti-alias
stopband rejection. Upgrade sum/diff lowpass filters from 4th-order to
6th-order Butterworth (three biquad stages) for ~36 dB/octave rolloff,
improving stereo separation by better rejecting the 38 kHz subcarrier
residuals.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
The fixed WFM_RESAMP_CUTOFF of 0.94 passed frequencies up to 94 kHz at
200 kHz composite rate, while the output Nyquist is only ~24 kHz. The
38 kHz demod products in the stereo diff path were only ~31 dB attenuated
by the Butterworth and aliased back into 10-20 kHz audio, causing treble
corruption in stereo mode. Now the cutoff is computed as
audio_rate / composite_rate, properly anti-aliasing the polyphase
resampler output.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
32 taps caused audio silence on real signals. Revert to 16 taps
which works correctly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Replace the low-accuracy 0.273 linear atan approximation with a
7th-order minimax polynomial (max error ~2.4e-7 rad vs ~0.004 rad).
Use branchless |y|>|x| argument reduction instead of y/x division
with quadrant fixup, avoiding division-by-zero and NaN branches.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Replace fast_atan2 polynomial approximation with f32::atan2 in the WFM
stereo decoder's FM discriminator and pilot PLL. The approximation
introduced harmonic distortion (~0.22 deg error) that manifested as
treble artifacts on strong/overdeviated broadcast signals.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Normalize IQ samples exceeding unit magnitude before the FM
discriminator. The discriminator only uses phase, so clamping
amplitude prevents overdeviated signals from producing clipped
composite baseband without losing frequency information.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Doubles the polyphase FIR length for the composite-to-audio rate
converter, improving stopband rejection from ~25 dB to ~50 dB.
This reduces treble distortion from imaging artifacts.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
- Set STEREO_DIFF_BW_HZ = AUDIO_BW_HZ so both filter paths have
identical group delay (improves multitone separation by ~10 dB).
- Zero out STEREO_SEPARATION_PHASE_TRIM (unnecessary with matched filters).
- Replace gradual blend ramp with binary blend: full stereo at pilot
lock, mono when unlocked. The hysteresis thresholds already handle
noisy signals.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Test L-only and R-only signals with tones at 400, 2000, 8000 and
14000 Hz to catch frequency-dependent group delay and phase trim
issues that the single 1 kHz test misses.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Restore AUDIO_BW_HZ to 15.8 kHz for cleaner mono path, widen
STEREO_DIFF_BW_HZ to 18 kHz for better high-frequency stereo detail,
and raise STEREO_MATRIX_GAIN from 0.30 to 0.50 (mathematically correct
unity gain for the L=(S+D)/2 stereo matrix).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Raise both AUDIO_BW_HZ and STEREO_DIFF_BW_HZ to 18 kHz so the L+R and
L-R filter paths have identical group delay across the full audio band.
The previous mismatch (15.8 vs 14.5 kHz) caused frequency-dependent
phase errors in the stereo matrix that degraded real-world separation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
Raise the stereo blend floor from 0.55 to 0.75 at pilot lock and lower
the full-blend ceiling from stereo_detect_level 0.92 to 0.70. This
gives real-world signals with moderate pilot strength much better L/R
separation (~17 dB immediately at lock vs ~5 dB before) and reaches
full unity blend sooner.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
sjg
