dashersw/coyopedal/measuring-pedal-audio

Measuring audio on the device

Core rule: diagnose from a fresh on-device measurement. Never reason from cycle or byte counts remembered from older builds, including the approximate figures in docs/. Placement changes move them, and stale numbers have already led to wrong conclusions.

Reaching the board

• The radios are off in audio mode. Wi-Fi and BLE run only in maintenance mode. To enter

it, tap the preset name, then Setup, then Maintenance mode. The BOOT button also works.

• Credentials live in src/native/services/remote_config.h (gitignored). Create them with

npm run remote:configure -- --ssid YOUR_WIFI.

• Remote tool: npm run remote -- [--host IP] <verb>. The main verbs are

discover, status, logs [--follow], command <TEXT>, ota <image>, audio-mode, live-status, effects-profile. command HELP lists every device command.

• The usual host is in .gea/boards.json (transports.ota.host).

The measurement loop

1. From maintenance, run command AUDIO TRY <5..300> [ENGAGE] [MASK <n>] [MODEL <idx>] [IR [idx]] [SIGNAL]. The

board boots into real audio for that many seconds, with radios off, then returns to maintenance. A timer enforces the return even if audio fails to start.

1. Wait for the board to come back, then run logs to read the reboot-surviving log ring.
2. Read the audio: heartbeat lines. audio_heartbeat_task in src/native/drivers/usb_audio.cpp

writes one every 5 s in every build:

audio: cap=+240024 play=+240024 silence=+0 rings=22/122 err=0 paused=0 bypass=0 amp=1
       in=0.04937 out=0.37613 ch=0.04937/0.00187
       blk=3750 C0=412/998us 31% C1=506/1120us 38% budget=1333us miss=0/0

MASK is a bitmask in coyopedal_fx_block_t order. This is the storage order, not the order the chain runs in: gate=1, compressor=2, overdrive=4, reverb=8, modulation=16, delay=32. For example, gate + compressor + reverb is MASK 11.

Reading it

bypass is the footswitch. amp is the amp block's own switch.

Other instruments

• HEAP MAP prints the maintenance heap census (format in docs/MEMORY.md, "How to

measure"); the audio-boot census is in logs.

• effects-profile --seconds N --input synthetic|live measures each effect alone, then the full chain.

Settled facts: don't re-test

The USB client task (prio 20) must preempt stage A. Split 8 beats split 9. Hot per-block code belongs in IRAM. RTC fast memory is slow for hot buffers. NOUI, CONFIG_ESP_PHY_IRAM_OPT, run_layer placement, model choice, SRAM bank layout and a heap-placed graph were all dead ends. docs/MEMORY.md ("Things that were tried and did not help") summarizes why. The interface on the bench is an XTONE Pro (UAC2, 24-bit packed stereo, no feedback endpoint), not an iRig.

Common mistakes

• Calling audio clean because the counters are clean. Only the user's ears can confirm that.
• Opening a USB serial monitor during a window. It can toggle reset or BOOT and invalidate the run.
• Measuring again just to confirm a result you already have. Measure only when the number changes what you do next.