Display Hardware — fb_ili79400 Circular LCD

Hardware

• Driver: fb_ili79400 (confirmed via /proc/fb)
• Framebuffer: 400×304 px, 16 bpp RGB565, stride 800 bytes
• Physical shape: circular LCD with a PWM backlight (/sys/class/leds/lcd-backlight, 0–255)
• Panel type confirmed on the device (2026-10-05): writing 0 to /sys/class/leds/lcd-backlight/brightness blanks the image completely. It is an LCD (the device tree lists an ilitek,ili79400 TDDI controller and a pwm-leds lcd-backlight), so there is no burn-in and brightness means backlight level

Visible Area (measured on the device, 2026-10-05)

The visible image is a circle of radius ≈198 px centred at ≈(200, 106) in framebuffer coordinates, cut off flat at the top by framebuffer row 0 (the area above it is glass under the stand’s notch, with no pixels). Left and right, the bezel sits just outside the framebuffer edges at the widest row; at the bottom the circle ends at about y 304.

Measured with calibration patterns written straight to /dev/fb0 (waved stopped) and straight-down photos: a circle of r 196 around (200, 105) hugs the bezel all the way round the curve, r 202 and r 208 fall outside it; the y = 0 row is visible from about x 20 to 380.

The earlier assumption here — a circle of r 190 centred at (200, 152) — was wrong by about 46 px vertically. Layouts built on it sit too low and leave the top third unused.

y	visible x range	width
0	33 → 367	335
32	16 → 384	369
64	7 → 393	387
106	2 → 398	397
160	10 → 390	381
220	38 → 362	325
260	76 → 324	249
290	127 → 273	147
303	180 → 220	41

The table is computed from the measured circle (r 198 at (200, 106)), clipped to the framebuffer. Keep tap targets ~8 px inside the curve.

flush() — Rotation + Byte-Swap

Two transforms must be applied together when writing LVGL pixels to /dev/fb0:

1. 180° rotation

The hardware is mounted upside-down. There is no HW rotation support (/sys/class/graphics/fb0/rotate = 0). Applied as vertical + horizontal flip in the flush callback.

2. Byte-swap per pixel

fb_ili79400 reads RGB565 pixels MSB-first (big-endian). ARM is little-endian. Each u16 pixel must be byte-swapped before writing.

// Sequential destination writes (cache-friendly), reversed source reads.
let dst_x_start = SCREEN_W - x2 - 1;
for row in 0..h {
    let phys_y = SCREEN_H - 1 - (y1 + row);
    let src = color_p.add(row * w * 2) as *const u16;
    let dst = fb_ptr.add((phys_y * SCREEN_W + dst_x_start) * 2) as *mut u16;
    for col in 0..w {
        *dst.add(col) = (*src.add(w - 1 - col)).swap_bytes();
    }
}

Why not lv_display_set_rotation(180)?

In PARTIAL render mode, LVGL transforms the lv_area_t coordinates passed to flush_cb into rotated space, but the pixel buffer remains in logical order. Writing logical-order pixels at physically-rotated coordinates scrambles colors. Manual rotation in flush() avoids this entirely.

LVGL Configuration (lv_conf.h)

Setting	Value	Reason
LV_COLOR_DEPTH	16	RGB565 matches fb0
LV_USE_STDLIB_MALLOC	LV_STDLIB_CLIB	System malloc required — decoded PNGs (~240 KB for 400×304) exceed the old 128 KB fixed pool
LV_USE_LODEPNG, LV_USE_FS_POSIX	0	No PNG assets since the milestone 2 runtime switch: icons come from the kit icon font
LV_FONT_DEFAULT	&font_body_15	The kit body font; Montserrat is not built

LV_COLOR_16_SWAP was removed in LVGL 9. Byte-swapping is handled in flush() instead.

These stock pspapp assets are no longer used by waved (milestone 2); the table stays as a record of the stock firmware.

Asset Paths on Device

Original pspapp assets at /etc/pspapp/lvgl/ (referenced via the A: LVGL FS drive):

Asset	Path	Size
Mode button background	A:/etc/pspapp/lvgl/icon/msc/bg_main_circule_btn.png	76×76
REC button background	A:/etc/pspapp/lvgl/icon/msc/bg_main_record_btn.png	180×76
Mute button background	A:/etc/pspapp/lvgl/icon/msc/bg_main_mic_btn.png	76×76
Beamforming mode icons	A:/etc/pspapp/lvgl/icon/beamforming/bf_mode_*.png	32×32
Battery status icons	A:/etc/pspapp/lvgl/icon/status/battery_*.png	20×20
WiFi status icons	A:/etc/pspapp/lvgl/icon/status/wifi_*.png	20×20
Bluetooth status icons	A:/etc/pspapp/lvgl/icon/status/bluetooth*.png	20×20
Mic icon	A:/etc/pspapp/lvgl/icon/volume/mic_on.png	—

Do NOT use bg_main.png — it is a 400×304 rectangular image with a glowing teal sphere that looks like a “weird orb” on the circular display.

Screen layout

The UI is the milestone-2 kit (crates/waved-ui/src/kit/README.md) on the approved mockups (../docs/mockups/). Geometry tokens live in kit/tokens.rs and follow the measured visible area above. The stock firmware’s own layout is kept for reference in docs/ui/HOME_SCREEN.md.

Boot screen

The Insta360 image at power-on is drawn by the bootloader and stays on the panel until something draws over it (fbtft keeps the panel’s contents; the kernel has no logo, CONFIG_LOGO is unset). It cannot be replaced at the source: the logo partition (/dev/logo, mmcblk0p5) is all zeros, the bootloader’s load_bmp_logo reads a missing odm_ext partition or that empty one and init_display never calls it, and the image lives in the bootloader itself, which is encrypted (entropy 8.0 bits/byte over 0xC0000–0x180000) and verified at boot (androidboot.verifiedbootstate=green).

deploy/S03splash covers it instead: right after S02overlayfs mounts the overlay (/etc/init.d and /data/waved both live on /data) it writes /data/waved/splash.fb (waved’s start-up ring without the version line, built by scripts/make-splash.py) to /dev/fb0. rcS sets the backlight to 255 before any script runs. Measured 2026-10-12: userspace starts at 0.97 s, the splash before S50 at 2.24 s, waved’s UI at ~10 s, so the Insta360 image now shows for the bootloader’s time plus ~1.5 s. Install: push both files and chmod 755 /etc/init.d/S03splash.

Deploy

/data is ~440 MB and nearly full; the stripped binary is ~63 MB. adb push over the running binary deletes the old file before it fails for space, leaving no binary. Deploy through /tmp and swap with the watchdog stopped.

deploy/install.sh does all of it, for a first install and for an update: after the cross build below, run it from anywhere. In this order, which is the safety net:

1. Checks, before anything changes: the Wave answers, no over-the-air update is on trial (/data/waved/update/trial, docs/UPDATES.md), /proc/mounts shows S02overlayfs’ layout (release mode: /etc and /usr overlays with upperdir=/data/overlay/upper/etc and …/usr; debug mode: one overlay on / with upperdir=/data/overlay/upper), and /data has room (df -Pk) for the stripped binary plus 20 MB, counting the installed binary’s space as free.
2. Strips the binary (llvm-objcopy from rustup component add llvm-tools, or OBJCOPY=...) and pushes it, S60waved, S03splash, rsyslog.conf, splash.fb, iptables.rules, scripts/patch-stock-usbgadget.sh and (if the Wave has none) config/default.toml to /tmp, checking each one’s md5.
3. Copies the Wave’s config.toml to /data/waved/config.toml.before-<UTC time> (mode 600); it never replaces it.
4. Runs the USB gadget patch (ADB at boot, docs/USB_STORAGE.md) and checks with awk that no adbd_enable in S89usbgadget is commented out and at least one is live inside start_usbgadget(). Nothing is stopped before this passes: a Wave without the stock app and without ADB can’t be reached again.
5. Stops the stock app and the watchdog, waits for waved to exit (killall -9 at 30 s, gives up at 40 s), swaps the binary (the old one waits in /tmp/waved.old, its md5 in /tmp/waved.old.md5, and goes back if the new one does not fit). The swap ignores SIGHUP/INT/TERM/PIPE: adbd SIGHUPs a session’s processes when it closes, and a swap cut half-way through the 63 MB copy would leave no binary.
6. Installs S60waved, S03splash, rsyslog.conf, splash.fb and iptables.rules, each through a dot-named copy and a rename; the default config.toml if there was none.
7. Last, the stubs that keep the stock app and its telemetry off (S60pspapp_daemon, libapplogrs.so). One already in /data/overlay/upper is kept as it is, whatever it holds (on some Waves libapplogrs.so there is the compiled no-op scripts/libapplogrs.so, 8352 bytes). A missing one is written through the /etc and /usr mounts and read back (#!/bin/sh and a newline; 0 or 1 byte). Then both must be in /data/overlay/upper: the copies there are what replace the stock files at boot.
8. An update starts the watchdog again, unless one runs already (rcS starts it without a pidfile; two would fight over the devices), and ends with an adb reboot hint when the gadget patch, S03splash, splash.fb or rsyslog.conf changed (by md5; they take effect at boot). A first install (no /data/waved/waved) asks for adb reboot. A failure, Ctrl-C, closed terminal (SIGHUP) or closed output pipe after step 5 began does the same; on an update it first puts /tmp/waved.old back if /data/waved/waved is missing or matches neither the new md5 nor the old one (it waits for a swap still running). The old md5 counts only if this run’s swap wrote it: the stop step deletes one an earlier update in the same boot left, and the swap writes it only after checking the copy.

Overlay files are written through the mounts, never into /data/overlay/upper directly (deploy/revert.sh deletes there, on purpose). It tests the Wave’s files on the Wave and checks every step’s output, because adb shell’s exit status is not reliable. Host-tested against a fake adb (deploy/tests/test-deploy-scripts.sh, crates/waved-update/src/install_script.rs).

By hand, the same steps for the binary alone:

S=$(mktemp -d)
WAVED_BUILD_ID=$(git rev-parse --short HEAD) \
  cross build --release --target aarch64-unknown-linux-gnu --features "ui,capture-alsa" -p waved
$(rustc --print sysroot)/lib/rustlib/aarch64-apple-darwin/bin/llvm-objcopy --strip-all \
  target/aarch64-unknown-linux-gnu/release/waved "$S/waved"
adb push "$S/waved" /tmp/waved.new && adb shell md5sum /tmp/waved.new   # compare with md5 -q "$S/waved"
adb shell 'kill $(ps | awk "/S60waved start/ && !/awk/ {print \$1}"); sleep 1; killall waved; \
  while pidof waved >/dev/null; do sleep 1; done; \
  rm /data/waved/waved && cp /tmp/waved.new /data/waved/waved && chmod 755 /data/waved/waved; \
  start-stop-daemon -S -b -m -p /tmp/S60waved.pid -x /bin/sh -- /etc/init.d/S60waved start'

• If ~/.cargo/config.toml sets rustc-wrapper = "sccache", run cross with a copy of Cross.toml (CROSS_CONFIG=...) that adds "RUSTC_WRAPPER=" to [build.env] passthrough. Don’t use CROSS_BUILD_ENV_PASSTHROUGH: it replaces the list and drops the CFLAGS ... -D__ARM_ARCH=8 that ring needs.
• Check adb shell df -h /data first. The stopped binary’s space is only freed once waved exits, which is why the swap happens with the watchdog and waved stopped. Shutdown takes ~2 s (the motor retracts first): if the cp starts before waved has exited it fails with “No space left on device” over a partial file. The new binary is still in /tmp, so wait for waved to exit and run the swap again.
• The init script SIGTERMs the first waved 5 s after it starts (display workaround) and rotates waved.log to waved.log.1 past 2 MB at each launch.
• Stock rsyslogd logged everything to /opt/logs/psp.log, which is on /data (overlay) and which nothing has rotated since the stock app stopped running. It reached 273 MB, mostly Wi-Fi driver (atbm_log) scan lines, and on 2026-10-08 it filled /data. Waved then exited at start-up without a word, because even its log write failed. deploy/rsyslog.conf drops the atbm_log lines and logs to /tmp/psp.log. install.sh installs it as /etc/rsyslog.conf (it lands in the overlay and survives reboots); it takes effect at the next boot, or at once with /etc/init.d/S11rsyslogd restart. If waved won’t start, check df -h /data first.
• start-stop-daemon needs -m -p <pidfile>; without it busybox matches any running /bin/sh and refuses to start.
• The framebuffer can be captured for checks with adb shell 'cat /dev/fb0 | head -c 243200' (RGB565 big-endian, rotated 180°; scripts/png-to-fb.py is the inverse).