[Problem]: Convolver reinitializes repeatedly during AirPlay 2 buffered playback

[cwawak]

What happened?

Hi, thanks all for all the awesome work.

I’m testing convolution support on a Raspberry Pi 4 plugged into a Yamaha amp. With AirPlay 2 Buffered playback from macOS (and iOS), Shairport repeatedly logs the same convolver being initialized during one playback session. I generated the FIR impulse response WAVs with the HouseCurve app on an iPhone.

With a unity FIR whose impulse is at sample 0, audio still plays. With a 0.5s centered FIR from HouseCurve, playback was silent or nearly silent.

I found a workaround, but I’m not sure it is the right fix.

The workaround was in player.c, inside player_thread_func around the convolver setup check. I tried removing inframe->length from the conditions that trigger convolver reinitialization

diff --git a/player.c b/player.c
@@ -4626,7 +4626,6 @@ void *player_thread_func(void *arg) {
                   // First, have we got the right convolution setup?

                   static int convolver_is_valid = 0;
-                  static size_t current_convolver_block_size = 0;
                   static unsigned int current_convolver_rate = 0;
                   static unsigned int current_convolver_channels = 0;
                   static double current_convolver_maximum_length_in_seconds = 0;
@@ -4635,7 +4634,6 @@ void *player_thread_func(void *arg) {
                     if (
                         // if any of these are true, we need to create a new convolver
                         // (conn->convolver_is_valid == 0) ||
-                        (current_convolver_block_size != inframe->length) ||
                         (current_convolver_rate != conn->input_rate) ||
                         !((current_convolver_channels == 1) ||
                           (current_convolver_channels == conn->input_num_channels)) ||
@@ -4646,7 +4644,6 @@ void *player_thread_func(void *arg) {
                       // look for a convolution ir file with a matching rate and channel count

                       convolver_is_valid = 0; // declare any current convolver as invalid
-                      current_convolver_block_size = inframe->length;
                       current_convolver_rate = conn->input_rate;
                       current_convolver_channels = conn->input_num_channels;

FFTConvolver.h says that after initialization, subsequent data portions of arbitrary length can be convolved.

Relevant log output

Before workaround:

Connection 1: AirPlay 2 Buffered playback. Source: "Christopher’s MacBook Pro". Input format: AAC/48000/F24/2. Output format: 44100/S32_LE/2.
player.c:4705 convolver: "/etc/shairport-sync/filters/unity-48000.wav".
player.c:4705 convolver: "/etc/shairport-sync/filters/unity-48000.wav".
player.c:4705 convolver: "/etc/shairport-sync/filters/unity-48000.wav".
... repeated continuously during playback ...

After workaround:

Connection 1: AirPlay 2 Buffered playback. Source: "Christopher’s MacBook Pro". Input format: AAC/48000/F24/2. Output format: 44100/S32_LE/2.
player.c:4702 convolver: "/etc/shairport-sync/filters/q100-cutsonly-48.wav".
Playback stopped. Total playing time 00:00:37. Output: 44099.49 (raw), 44099.75 (corrected) frames per second.

System Information.

Raspberry Pi 4, Debian 12 bookworm, Yamaha USB DAC via ALSA

Configuration Information.

0.006827863 "shairport.c:1730" config.service_type is: "auto".
         0.000930701 "shairport.c:1778" NQPTP service found.
         0.000074666 "shairport.c:2523" >> Display Config Start.
         0.003175896 "shairport.c:2523"
         0.000058685 "shairport.c:2523" From "uname -a":
         0.000011555 "shairport.c:2523"  Linux nad3020streamer 6.6.70-v8+ #1837 SMP PREEMPT Fri Jan 10 12:02:51 GMT 2025 aarch64 GNU/Linux
         0.007090104 "shairport.c:2523"
         0.000394258 "shairport.c:2523" From /etc/os-release:
         0.000171036 "shairport.c:2523"  Debian GNU/Linux 12 (bookworm)
         0.003575024 "shairport.c:2523"
         0.000310943 "shairport.c:2523" From /sys/firmware/devicetree/base/model:
         0.000044963 "shairport.c:2523"  Raspberry Pi 4 Model B Rev 1.4
         0.000173055 "shairport.c:2523"
         0.000035481 "shairport.c:2523" Shairport Sync Version String:
         0.000031111 "shairport.c:2523"  5.1-dirty-AirPlay2-smi10-OpenSSL-Avahi-ALSA-soxr-convolution-sysconfdir:/etc
         0.000224944 "shairport.c:2523"
         0.000038963 "shairport.c:2523" Command Line:
         0.000082185 "shairport.c:2523"  shairport-sync --displayConfig
         0.000074518 "shairport.c:2523"
         0.000034389 "shairport.c:2523" Configuration File:
         0.000030277 "shairport.c:2523"  /etc/shairport-sync.conf
         0.000029871 "shairport.c:2523"
         0.000467331 "shairport.c:2523" Configuration File Settings:
         0.000046352 "shairport.c:2523"  general :
         0.000031296 "shairport.c:2523"  {
         0.000029352 "shairport.c:2523"    name = "Office Receiver";
         0.000029296 "shairport.c:2523"    volume_range_db = 35;
         0.000030000 "shairport.c:2523"    volume_control_profile = "standard";
         0.000029629 "shairport.c:2523"  };
         0.000028834 "shairport.c:2523"  alsa :
         0.000203073 "shairport.c:2523"  {
         0.000091926 "shairport.c:2523"    output_device = "hw:DAC";
         0.000033888 "shairport.c:2523"    output_rate = 44100;
         0.000029926 "shairport.c:2523"  };
         0.000024389 "shairport.c:2523"  pa :
         0.000010778 "shairport.c:2523"  {
         0.000008240 "shairport.c:2523"  };
         0.000008056 "shairport.c:2523"  dsp :
         0.000007963 "shairport.c:2523"  {
         0.000007907 "shairport.c:2523"    convolution_enabled = "yes";
         0.000008278 "shairport.c:2523"    convolution_thread_pool_size = 1;
         0.000008611 "shairport.c:2523"    convolution_ir_files = "/etc/shairport-sync/filters/q100-cutsonly-44.wav, /etc/shairport-sync/filters/q100-cutsonly-48.wav";
         0.000009482 "shairport.c:2523"    convolution_gain = -4.0;
         0.000008759 "shairport.c:2523"    convolution_max_length_in_seconds = 0.5;
         0.000008500 "shairport.c:2523"    loudness_enabled = "no";
         0.000007907 "shairport.c:2523"  };
         0.000007870 "shairport.c:2523"  diagnostics :
         0.000007982 "shairport.c:2523"  {
         0.000007833 "shairport.c:2523"    statistics = "yes";
         0.000008056 "shairport.c:2523"    log_verbosity = 1;
         0.000008222 "shairport.c:2523"    log_show_time_since_last_message = "no";
         0.000008815 "shairport.c:2523"  };
         0.000008222 "shairport.c:2523"
         0.000007926 "shairport.c:2523" >> Display Config End.
         0.000008611 "shairport.c:2525" >> Goodbye!

PulseAudio or PipeWire installed?

• Check if your system uses a Sound Server.

How did you install Shairport Sync?

Built from source

Check previous issues

• Confirm

[mikebrady]

Thanks for the interesting report! Would you be kind enough to post the convolver files that are involved, please?

[cwawak]

Of course, sorry for the delay! I’ve attached the two exact 48 kHz IR files involved:

• unity-48000.wav: impulse at sample 0. The convolver repeatedly reinitialized, but audio remained audible.
• q100-cutsonly-48.wav: HouseCurve FIR centered at sample 12000 (250 ms). Reinitialization made playback silent. It played normally with the workaround.

Both are mono, 32-bit floating-point WAVs, 24,000 samples long.

shairport-convolver-issue-2236.zip

[mikebrady]

Thanks for these. I could run them both without any particular problems, but on a fast machine -- at present I don't have access to a Pi 4. Here is a sample log. It's from the development version of Shairport Sync, a little ahead of the next public release, but nothing has been done to the convolution stuff. It's also on FreeBSD, but that likewise should not be relevant.

$ shairport-sync -v --statistics
         0.000246250 "../shairport.c:631" looking for configuration file at full path "/usr/local/etc/shairport-sync.conf"
         0.002978750 "../shairport.c:1751" config.service_type is: "auto".
         0.000690042 "../shairport.c:1799" NQPTP service found.
         0.000228875 "../shairport.c:2696" Startup in AirPlay 2 mode, with features 0x401c4a00,0x50340 on device "00:0c:29:9f:c9:7e".
         0.000156250 "../shairport.c:2737" Version String: "5.2-dev-43-g96367998-dirty-AirPlay2-smi10-OpenSSL-Avahi-sndio-convolution-metadata-mqtt-dbus-sysconfdir:/usr/local/etc"
         0.000065417 "../shairport.c:2756" Command Line: "./shairport-sync -v --statistics".
         0.000237333 "../shairport.c:2800" audio backend is "sndio".
         0.035789375 "../shairport.c:3091" disable_resend_requests is off.
         0.000113458 "../shairport.c:3095" diagnostic_drop_packet_fraction is 0.000000. A value of 0.0 means no packets will be dropped deliberately.
         0.000034875 "../shairport.c:3096" statistics_requester status is 1.
         0.000030542 "../shairport.c:3101" rtsp listening port is 7000.
         0.000031667 "../shairport.c:3102" udp base port is 6001.
         0.000154833 "../shairport.c:3103" udp port range is 10.
         0.000039958 "../shairport.c:3104" player name is "FreeBSD15x".
         0.000033334 "../shairport.c:3106" run_this_before_play_begins action is "(null)".
         0.000075000 "../shairport.c:3107" run_this_after_play_ends action is "(null)".
         0.000068208 "../shairport.c:3108" wait-cmd status is 0.
         0.000040083 "../shairport.c:3110" run_this_before_play_begins may return output is 0.
         0.000104084 "../shairport.c:3112" run_this_if_an_unfixable_error_is_detected action is "(null)".
         0.000031916 "../shairport.c:3114" run_this_before_entering_active_state action is  "(null)".
         0.000030709 "../shairport.c:3116" run_this_after_exiting_active_state action is  "(null)".
         0.000048750 "../shairport.c:3117" active_state_timeout is  10.000000 seconds.
         0.000050333 "../shairport.c:3118" mdns backend "(null)".
         0.000073542 "../shairport.c:3124" interpolation setting is "vernier".
         0.000054041 "../shairport.c:3126" interpolation soxr_delay_threshold is 30000000.
         0.000050084 "../shairport.c:3127" resync time is 0.050000 seconds.
         0.000031541 "../shairport.c:3129" allow a classic AirPlay session to be interrupted: "no".
         0.000052459 "../shairport.c:3130" busy timeout time is 60.
         0.000044916 "../shairport.c:3131" drift tolerance is 0.002000 seconds.
         0.000049250 "../shairport.c:3132" password is "(null)".
         0.000032959 "../shairport.c:3133" default airplay volume is: -24.000000.
         0.000036583 "../shairport.c:3134" ignore_volume_control is 0.
         0.000043750 "../shairport.c:3138" volume_max_db is not set
         0.000106792 "../shairport.c:3141" volume range in dB (zero means use the range specified by the mixer): 0.
         0.000037541 "../shairport.c:3145" volume_range_combined_hardware_priority (1 means hardware mixer attenuation is used first) is 0.
         0.000032750 "../shairport.c:3148" playback_mode is 0 (0-stereo, 1-mono, 1-reverse_stereo, 2-both_left, 3-both_right).
         0.000050500 "../shairport.c:3149" disable_synchronization is 0.
         0.000044917 "../shairport.c:3150" use_mmap_if_available is 1.
         0.000080333 "../shairport.c:3152" output_format automatic selection is enabled.
         0.000044667 "../shairport.c:3157" output_rate automatic selection is enabled.
         0.000033292 "../shairport.c:3161" audio backend desired buffer length is 1.000000 seconds.
         0.000044375 "../shairport.c:3164" audio_backend_buffer_interpolation_threshold_in_seconds is 0.250000 seconds.
         0.000043875 "../shairport.c:3166" audio backend latency offset is 0.000000 seconds.
         0.000051500 "../shairport.c:3168" audio backend silence lead-in time is "auto".
         0.000048500 "../shairport.c:3172" zeroconf regtype is "_raop._tcp".
         0.000045625 "../shairport.c:3175" decoders_supported bit field is 4 (1 == hammerton, 2 == apple, 4 == ffmpeg).
         0.000048666 "../shairport.c:3176" decoder_in_use is 4.
         0.000062209 "../shairport.c:3177" alsa_use_hardware_mute is 0.
         0.000030458 "../shairport.c:3181" no special mdns service interface was requested.
         0.000050375 "../shairport.c:3185" configuration file name "/usr/local/etc/shairport-sync.conf" resolves to "/usr/local/etc/shairport-sync.conf".
         0.000050667 "../shairport.c:3192" metadata enabled is 1.
         0.000036875 "../shairport.c:3193" metadata pipename is "/tmp/shairport-sync-metadata".
         0.000120250 "../shairport.c:3195" metadata socket address is "(null)" port 0.
         0.000039416 "../shairport.c:3197" metadata socket packet size is "500".
         0.000031084 "../shairport.c:3198" get-coverart is 1.
         0.000048750 "../shairport.c:3201" mqtt is disabled.
         0.000062208 "../shairport.c:3203" mqtt hostname is (null), port is 1883.
         0.000032167 "../shairport.c:3204" mqtt topic is /FreeBSD15x.
         0.000051541 "../shairport.c:3206" mqtt will not publish raw metadata.
         0.000035625 "../shairport.c:3208" mqtt will not publish parsed metadata.
         0.000044125 "../shairport.c:3210" mqtt will not publish cover Art.
         0.000050375 "../shairport.c:3212" mqtt will not set retain flag.
         0.000093084 "../shairport.c:3214" mqtt remote control is disabled.
         0.000031333 "../shairport.c:3216" mqtt autodiscovery is disabled.
         0.000046042 "../shairport.c:3221" convolution_enabled is true.
         0.000046458 "../shairport.c:3223" convolution maximum length is 1.000000 seconds.
         0.000043750 "../shairport.c:3224" convolution gain is -4.000000
         0.000044833 "../common.c:2574" convolution impulse response files: 1 found.
         0.000255084 "../common.c:2589" convolution impulse response file "/usr/local/etc/shairport-sync/q100-cutsonly-48.wav": 24000 frames (0.5 seconds), 1 channel at 48000 frames per second.
         0.000065375 "../shairport.c:3229" loudness_enabled is false.
         0.000063458 "../shairport.c:3230" loudness reference level is -16.000000
         0.000379792 "../FFTConvolver/convolver.cpp:50" thread pool initialized with 1 thread and 8 convolvers.
         0.000199666 "../dbus-service.c:1293" Looking for a Shairport Sync native D-Bus interface "org.gnome.ShairportSync" on the system bus.
         0.003209000 "../dbus-service.c:1274" *warning: could not acquire a Shairport Sync native D-Bus interface "org.gnome.ShairportSync" on the system bus.
        38.150456333 "../common.c:2387" Can not set realtime properties of thread "ap2_ba_1".
         0.712873084 "../player.c:1007" Channel Mapping: | 0 ("FL") <- FL | 1 ("FR") <- FR |
         0.000537208 "../player.c:4209" Connection 1: AirPlay 2 Buffered playback. Source: "Mike Brady's iPhone". Input format: AAC/48000/F24/2. Output format: 48000/S16_LE/2.
         0.002679833 "../player.c:4706" convolver: "/usr/local/etc/shairport-sync/q100-cutsonly-48.wav".
         7.921938959 "../player.c:4130" Av Sync Error (ms) | Net Sync PPM | All Sync PPM | Av Sync Window (ms) | Min DAC Queue | Min Buffers | Min Buffer Size
         0.000039666 "../player.c:4130"             -12.12         1481.6         1897.5                 21.76             928            27              230k
         8.021283084 "../player.c:4130"              -1.25          588.2          817.2                  3.40            4205            30             2678k

A couple of points.

1. You can see that Shairport Sync was able to read the convolution file:

            0.000255084 "../common.c:2589" convolution impulse response file "/usr/local/etc/shairport-sync/q100-cutsonly-48.wav": 24000 frames (0.5 seconds), 1 channel at 48000 frames per second.
   

2. The settings are basic:

   $ shairport-sync --displayConfig
   >> Display Config Start.
   
   From "uname -a":
    FreeBSD FreeBSD15x 15.1-RELEASE-p1 FreeBSD 15.1-RELEASE-p1 releng/15.1-n283582-0f691888dc56 GENERIC arm64
   
   From /etc/os-release:
    FreeBSD 15.1-RELEASE-p1
   
   Shairport Sync Version String:
    5.2-dev-43-g96367998-dirty-AirPlay2-smi10-OpenSSL-Avahi-sndio-convolution-metadata-mqtt-dbus-sysconfdir:/usr/local/etc
   
   Command Line:
    ./shairport-sync --displayConfig
   
   Configuration File:
    /usr/local/etc/shairport-sync.conf
   
   Configuration File Settings:
    dsp : 
    {
      convolution_enabled = "yes";
      convolution_ir_files = "/usr/local/etc/shairport-sync/q100-cutsonly-48.wav";
    };
   
   >> Display Config End.
   >> Goodbye!
   

So, I wonder if it is just that you need to allow the convolver to use more parallelism, i.e. more threads. Let me suggest that you change the convolution_thread_pool_size to, say, 3:

convolution_thread_pool_size = 3;

(Maybe you've tried this -- apologies if so...)

My experience is that running the convolver on multiple threads on a Pi 3 was really quite noisy -- due, I believe, to power supply transients -- but a Pi 4 was much better and a Pi 5 was fine.

Also, I notice that you are setting the output to 44,100 frames per second. If your DAC can accommodate it, 48,000 would be easier on the Pi, as transcoding would not be needed.

[mikebrady]

Just playing a bit more with this -- the problem arises in setting the output rate to 44100.

Pretty sure this is due to a mixup in the input rate (e.g. 48,000), the output rate (44,100 in this case), and the rate at which convolution is done. Let us try to figure this out...

[mikebrady]

Hi there. We just pushed an update into the development branch for your consideration. If you got a chance to test it out and report back, it would be great.

The problem was caused by (us!) confusing input and output rates and channel counts.

TL;DR: Convolution runs on the output, not the input, and thus must match the output rate (and, optionally, the output channel count). We confused these in the code and this caused the problems you encountered.

So, the FIR file has to match the output rate rather than the input rate. Thus, you'll need to create the FIR files at 44,100 fps.

[mikebrady]

Update -- a few extra changes were made; please make sure it's version 5.2-dev-42-g5e48d8fe or later.

[cwawak]

Tested and working. Thank you!

I built development at 081aa0ca (5.2-dev-43-g081aa0ca, which includes 5e48d8fe) on the Pi 4. (Built from a git worktree, so -V reports plain 5.2-dev the .git/index check in configure.ac doesn't trigger there.)

With output_rate = 44100 and a 44.1 kHz FIR, AirPlay 2 Buffered AAC/48000/F24/2 input: the correct FIR is selected exactly once, room-corrected audio plays, and it stayed stable through pause/resume, track changes, and a session teardown/replacement. Same result at output_rate = 48000 with the 48 kHz FIR. Sync stats are clean (~0.1–0.6 ms average error, no stuffing, thread pool still at 1).

For comparison, I had also tested 2346cd7a (before 5e48d8fe): with 44.1 kHz output it selected the correct FIR but reinitialized continuously 1,338 times over ~74 s, a steady 18–19/s with no audible music. That rate matches the resampler arithmetic: 48 kHz input to 44.1 kHz output yields 1024 × 44100/48000 = 940.8 frames per packet, so swr_convert emits 941 frames on most calls and 940 on the rest (two length transitions per five packets predicts ~18.75/s), and each transition tripped the current_convolver_block_size != inframe->length check you've now disabled.

On whether the disable can be permanent, two observations from the bundled FFTConvolver: init rounds the block size up with NextPowerOf2, so 940 and 941 built identical 1024-sample partition structures; and each init discards accumulated history, which is why a unity impulse at sample 0 stayed audible under repeated reinitialization while a centered FIR went silent.

Happy to test anything further, I'm learning a lot!

[mikebrady]

Excellent! Thanks for the very informative update -- I am likewise learning a lot and delighted to get such a thorough checking out!

I've just set the block size to 1,024 and pushed the update -- 5.2-dev-44-geb49aaab. We'll see (next month, unfortunately) how that does on low-powered devices.