Monorepo for Aesthetic.Computer
aesthetic.computer
native: negotiate S32_LE format for SOF speaker PCM (fixes crunchy/quiet)
SOF topology FE PCMs use S32_LE internally and the SSP1 BE DAI
(MAX98360A) runs S24_LE. Writing S16_LE samples to hw:0,0 caused
48dB attenuation + quantization noise because the DSP interpreted
16-bit values as 32-bit.
Try S32_LE format first in hw_params. If accepted, the audio thread
widens int16 samples to int32 (left-shift 16) before snd_pcm_writei.
Falls back to S16_LE for non-SOF hardware. Reverts to hw:0,0 as
primary device (plughw: broke audio entirely on SOF).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+42
-8
+41
-8
fedac/native/src/audio.c
+41
-8
fedac/native/src/audio.c
···
1124
1124
ACAudio *audio = (ACAudio *)arg;
1125
1125
const unsigned int period_frames = audio->actual_period ? audio->actual_period : AUDIO_PERIOD_SIZE;
1126
1126
int16_t *buffer = calloc(period_frames * AUDIO_CHANNELS, sizeof(int16_t));
1127
-
if (!buffer) { fprintf(stderr, "[audio] thread: alloc failed\n"); return NULL; }
1127
+
int32_t *buffer32 = NULL;
1128
+
if (audio->use_s32)
1129
+
buffer32 = calloc(period_frames * AUDIO_CHANNELS, sizeof(int32_t));
1130
+
if (!buffer || (audio->use_s32 && !buffer32)) {
1131
+
fprintf(stderr, "[audio] thread: alloc failed\n"); return NULL;
1132
+
}
1128
1133
const double rate = (double)(audio->actual_rate ? audio->actual_rate : AUDIO_SAMPLE_RATE);
1129
1134
const double dt = 1.0 / rate;
1130
1135
double mix_divisor = 1.0; // Smooth auto-mix (matches speaker.mjs)
···
1588
1593
rem2 -= f2; off2 += f2;
1589
1594
}
1590
1595
}
1596
+
/* Widen int16→int32 for S32_LE PCMs (SOF topology).
1597
+
* Left-shift by 16 places the 16-bit sample in the top half
1598
+
* of the 32-bit word, which maps correctly to the S24_LE BE
1599
+
* DAI after the DSP truncates the bottom 8 bits. */
1600
+
const void *write_buf = buffer;
1601
+
if (buffer32) {
1602
+
for (int j = 0; j < (int)(period_frames * AUDIO_CHANNELS); j++)
1603
+
buffer32[j] = (int32_t)buffer[j] << 16;
1604
+
write_buf = buffer32;
1605
+
}
1591
1606
int remaining = (int)period_frames;
1592
1607
int offset = 0;
1593
1608
while (remaining > 0) {
1594
-
int frames = snd_pcm_writei(pcm, buffer + offset * AUDIO_CHANNELS, remaining);
1609
+
const void *wptr = buffer32
1610
+
? (const void *)(buffer32 + offset * AUDIO_CHANNELS)
1611
+
: (const void *)(buffer + offset * AUDIO_CHANNELS);
1612
+
int frames = snd_pcm_writei(pcm, wptr, remaining);
1595
1613
if (frames == -EAGAIN) continue;
1596
1614
if (frames < 0) {
1597
1615
int rec = snd_pcm_recover(pcm, frames, 1);
···
1621
1639
}
1622
1640
1623
1641
free(buffer);
1642
+
free(buffer32);
1624
1643
return NULL;
1625
1644
}
1626
1645
···
1899
1918
snprintf(ucm_speaker_plug, sizeof(ucm_speaker_plug),
1900
1919
"plughw%s", ucm_speaker_pcm + 2); /* hw:0,0 → plughw:0,0 */
1901
1920
int n = 0;
1902
-
/* Prefer plug-wrapped device FIRST — the plug layer handles
1903
-
* format/rate/channel conversion, which fixes the "crunchy
1904
-
* quiet" audio on SOF boards where the SSP1 BE DAI runs at
1905
-
* a different format than our S16_LE 48kHz stereo request. */
1921
+
/* Prefer raw hw: first — SOF topology FE PCM runs at S32_LE
1922
+
* internally, and we now negotiate S32_LE directly so the DSP
1923
+
* does zero conversion. plughw: as fallback if hw: fails. */
1924
+
devices_with_spk[n++] = ucm_speaker_pcm;
1906
1925
devices_with_spk[n++] = ucm_speaker_plug;
1907
-
devices_with_spk[n++] = ucm_speaker_pcm;
1908
1926
for (int i = 0; devices_default[i] && n < 15; i++)
1909
1927
devices_with_spk[n++] = devices_default[i];
1910
1928
devices_with_spk[n] = NULL;
···
1968
1986
snd_pcm_hw_params_alloca(¶ms);
1969
1987
snd_pcm_hw_params_any(pcm, params);
1970
1988
snd_pcm_hw_params_set_access(pcm, params, SND_PCM_ACCESS_RW_INTERLEAVED);
1971
-
snd_pcm_hw_params_set_format(pcm, params, SND_PCM_FORMAT_S16_LE);
1989
+
/* SOF topology FE PCMs use S32_LE internally; the SSP1 BE DAI
1990
+
* (MAX98360A) runs S24_LE. Writing S16_LE to this pipeline
1991
+
* causes 48dB attenuation + quantization noise ("crunchy quiet").
1992
+
* Try S32_LE first — if the FE PCM accepts it, we write int32
1993
+
* samples and the DSP does zero conversion. Fall back to S16_LE
1994
+
* for non-SOF hardware (HDA, USB, etc). */
1995
+
audio->use_s32 = 0;
1996
+
if (snd_pcm_hw_params_set_format(pcm, params, SND_PCM_FORMAT_S32_LE) == 0) {
1997
+
audio->use_s32 = 1;
1998
+
fprintf(stderr, "[audio] Negotiated S32_LE format\n");
1999
+
} else {
2000
+
snd_pcm_hw_params_any(pcm, params);
2001
+
snd_pcm_hw_params_set_access(pcm, params, SND_PCM_ACCESS_RW_INTERLEAVED);
2002
+
snd_pcm_hw_params_set_format(pcm, params, SND_PCM_FORMAT_S16_LE);
2003
+
fprintf(stderr, "[audio] Negotiated S16_LE format (S32_LE not supported)\n");
2004
+
}
1972
2005
snd_pcm_hw_params_set_channels(pcm, params, AUDIO_CHANNELS);
1973
2006
1974
2007
// Query hardware rate range
+1
fedac/native/src/audio.h
+1
fedac/native/src/audio.h
···
244
244
int card_index; // ALSA card number (0 or 1)
245
245
unsigned int actual_rate; // Negotiated ALSA sample rate (may differ from requested)
246
246
unsigned int actual_period; // Negotiated ALSA period size in frames
247
+
int use_s32; // 1 if PCM negotiated S32_LE (SOF boards), 0 for S16_LE
247
248
248
249
// TTS PCM buffer (resampled to output rate, mono → stereo in mix)
249
250
float *tts_buf; // ring buffer of mono float samples at output rate