H700 port 03 Platform layer

03 — Platform Layer (workspace/h700/)

Reference map for the platform API, device detection, input architecture, settings backend, and H700-specific tools. Evidence is indexed in 08 and release classification in 09.

Strategy, as planned and executed: clone workspace/tg5040/, keep the generic_video.c/generic_wifi.c/generic_bt.c includes and the current PLAT_* API surface, repoint hardware access to the H700 paths (00), borrowing only hardware constants from the old rg35xxplus port (git show 8cd78866:...).

One platform serves all devices: h700-device.sh scrapes stock RGXX_MODEL and exports exact DEVICE SKU tokens (see 02). detect_device() keys only on DEVICE and sets panel_w / panel_h (app framebuffer size) plus stick/LED capability flags. Panel rotation on RG28XX is handled entirely by the SDL driver via SDL_ROTATION=1 (needs_portrait_sdl); nothing app-side rotates — 04.

Analog stick / LED matrix (DEVICE contract):

$DEVICE	Sticks	RGB LEDs
rg28xx, rg34xx, rgsp, rg35xxplus, rg35xxsp	none	no
rg34xxsp, rg35xxh, rg35xxpro	dual	no
rg40xxv	left only	1 bank
rg40xxh, rgcubexx	dual	2 banks

Unknown models map to rg35xxplus (640×480, no sticks, no RGB). No H700 Anbernic device has an Fn switch. The H700 PS1 pak binds L3/R3 so DualShock titles can use stick clicks on models that have them.

The RG SP is stickless on device-tree evidence rather than spec sheets: its DTB has no keyL3/keyR3 nodes, none of the analog multiplexer pins the RG34XXSP carries (amux-en-gpios, A0_gpio, A1_gpio, adc-en-gpios), and the GPADC itself is status = "disabled".

The Input pak uses the per-device capability flags to show only the sticks that exist, including their analog axis movement and L3/R3 click state. This is verified on the single-stick RG40XXV and dual-stick RG34XXSP: movement is visualized correctly for the left stick on RG40XXV and for both sticks on RG34XXSP.

platform.h (as shipped)

extern int panel_w, panel_h;       // app framebuffer from DEVICE
extern int needs_portrait_sdl;     // DEVICE=rg28xx
extern int hdmi_active;
extern int dev_has_lstick, dev_has_rstick;
extern int dev_has_rgb, dev_num_leds;

#define HDMI_WIDTH   1280
#define HDMI_HEIGHT  720
#define FIXED_SCALE   2
#define FIXED_WIDTH   (hdmi_active?HDMI_WIDTH:panel_w)
#define FIXED_HEIGHT  (hdmi_active?HDMI_HEIGHT:panel_h)
#define FIXED_BPP     2
#define SCREEN_FPS    60.0            // ⚠ assumed, never measured per panel
#define SDCARD_PATH   "/mnt/SDCARD"   // symlink to the real TF2 mountpoint (02)
#define MAX_LIGHTS    2               // ceiling; PLAT_getNumLeds() reports 0/1/2 per model
#define MAIN_ROW_COUNT ((hdmi_active||panel_h>=720)?10:6)
#define PADDING        ((hdmi_active||panel_h>=720)?5:10)

Input — dual path, evdev primary (deviation from plan)

The plan preferred the tg5040-style pure-SDL-joystick route. Shipped: raw evdev is the primary path for built-in controls, with SDL joystick as a secondary path for external controllers. Reasons discovered during bring-up:

• SDL is built without udev and SDL_JOYSTICK_DISABLE_UDEV=1 is exported. During bring-up the built-in pad did not enumerate because the SDL fork's Batocera patches deleted the joystick heuristic in SDL_EVDEV_GuessDeviceClass(), so the no-udev fallback path never classified any device as a joystick. Fixed by h700-toolchain's support/sdl2-h700.patch (see 01); SDL joystick enumeration of the built-in pad now works. Note the fork assigns SDL button indices in ascending evdev-keycode order, so the pad's ESC/VOL−/VOL+ (1/114/115) occupy indices 0–2 and the gamepad cluster starts at index 3 (A=3 … MENU=11). Because of that, platform.c must never open the built-in pad as an SDL joystick: poll_sdl_input() interprets SDL joysticks with the BT-pad JOY_* layout, so the built-in pad's events would double-apply with scrambled meanings (B→L1, VOL−→back, MENU→volume UI — seen in Settings 2026-07-09). is_builtin_pad() skips "ANBERNIC-keys" in both the init scan and the hotplug path; evdev is the sole path for built-in controls, SDL joystick is for external (BT) pads only.
• evdev codes are needed for keymon and wake handling anyway.

Implementation: poll_evdev_input() scans/reads /dev/input/event0..11 directly with periodic (2 s) rescans for hotplug, mapping via CODE_* defines; poll_sdl_input() handles SDL joystick events (BT pads) via JOY_* indices.

Verified mappings (also recorded in 00):

• evdev (event1 ANBERNIC-keys): A 304, B 305, Y 306, X 307, L1 308, R1 309, SELECT 310, START 311, MENU 312, L3 313, L2 314, R2 315, R3 316; PLUS 115 / MINUS 114 (event2); POWER 116 (event0, axp2202-pek — note tg5040 uses 102). D-pad has no keycodes on current firmware — it arrives as ABS_HAT0X/Y.
• MENU compound-tap quirk (fixed): the firmware reports the physical MENU button faithfully on 312 (down while held, up on release), but on a short tap it also emits a synthetic 354 (KEY_GOTO) pulse that starts the instant 312 releases and lasts ~190 ms (verified via evtest on RG34XXSP; on a long hold 354 never fires). Mapping both 312 and 354 to BTN_MENU (as the old rg35xxplus platform did via CODE_MENU_ALT) stretched every tap past the 250 ms MENU_DELAY threshold — a quick MENU tap registered as a hold, so the home screen flipped from the shortcuts overlay into brightness mode. Fix: button_from_code() no longer maps CODE_MENU_ALT to BTN_MENU (keymon likewise ignores it); tap-vs-hold is derived from the clean 312 timing. The SDL path was never affected (JOY_MENU_ALT = JOY_NA).
• Analog sticks: ABS_Z/RX/RY/RZ; signed values are scaled ×32767/4096. Presence varies per model (see the stick matrix above); dev_has_lstick/dev_has_rstick gate CODE_L3/R3, JOY_L3/R3, and AXIS_LX/LY/RX/RY to NA where the stick is absent, so the Input pak and shared input code adapt without platform-independent changes. (Previously CODE_L3/R3 were unconditional, so the Input pak drew L3/R3 pills on every device.)
• SDL indices (secondary path): A=0 B=1 Y=2 X=3 L1=4 R1=5 SELECT=6 START=7 MENU=8, L3=9 L2=10 R2=11 R3=12, MINUS=15 PLUS=16; axes LX=0 LY=1 RX=2 RY=3.

Button semantics: BTN_RESUME=BTN_X, BTN_SLEEP/WAKE=BTN_POWER, BTN_MOD_BRIGHTNESS=BTN_MENU, BTN_MOD_COLORTEMP=BTN_SELECT, dedicated BTN_MOD_PLUS/MINUS on the volume keys — parity with tg5040 conventions.

Not supported: user-assignable FN1/FN2/HOME pak-launch actions (upstream #788). H700 has no dedicated FN1/FN2/HOME buttons, so BTN_FN* are BTN_NONE with empty names (Settings hides the Assignments menu; main-menu FN presses are no-ops).

PLAT_shouldWake keeps a persistent wake_fd on event0 (opened once O_NONBLOCK|O_CLOEXEC, drained per poll, closed in PLAT_quitInput) — an early version open/closed per poll and could drop the wake press between polls.

platform.c — subsystem map (as shipped)

Subsystem	Implementation
Video	#include "generic_video.c" — custom SDL2 mali driver does the rest (04)
Battery	axp2202-battery/capacity + axp2202-usb/online, with coarse bucketing in shared code. Unlike tg5040, H700 charging detection uses online alone
CPU speed	governor.sh via system(): auto=schedutil and performance both permit the greatest advertised frequency; powersave=conservative capped mid-range. H700's advertised 1.5 GHz ceiling is in-spec, not an overclock, so Auto no longer caps it one step below maximum. Manual changes work. Former MD Auto slowdown near 480 MHz is fixed (user-verified 2026-07-20). Single A53 cluster → PLAT_pinToCores no-op
CPU temp	thermal_zone0
GPU temp	thermal_zone1 (zone map in 00 — zone2 is the video engine, a first draft got this wrong)
GPU speed / governor	HUD read resolves devfreq/clk paths once (first readable candidate wins) with a 648 MHz last-resort fallback. governor.sh also pins GPU min_freq to max under auto/performance and to min under powersave — simple_ondemand under-clocks vsync'd spiky loads (06)
Rumble	echo 1/0 > axp2202-battery/moto — on/off only, strength>0 → 1. Works (tested). Input-FF (event1 advertises FF bits) unexplored
LEDs	RGB on RG40XX H / RG40XX V / RG CubeXX only, via an MCU on /dev/ttyS5 (115200 8N1) gated by axp2202-battery/mcu_pwr — see platform/led.c. MAX_LIGHTS 2 is a ceiling; PLAT_getNumLeds() returns 0 (no RGB), 1 (RG40XXV, one populated bank) or 2. Effects are rendered by the MCU firmware, not in userspace. work_led remains the sleep/backlight indicator on every model
Backlight	raw brightness 0 via disp ioctl + fb blank + work_led on/off around it
Lid	hallkey path wired into PLAT_initLid/PLAT_lidChanged (has_lid = file exists); lid close → sleep and lid open wakes screen-off. RG34XXSP: POWER can wake light sleep with lid closed; deep sleep re-sleeps if the lid remains closed (06).
Model	PLAT_getModel copies the raw RGXX_MODEL into a static buffer; when unavailable it returns Anbernic RG XX
Date/time	timedatectl / hwclock / date via snprintf-bounded commands; timezone choices are parsed from /usr/share/zoneinfo/zone.tab, setting uses timedatectl set-timezone, and NTP uses set-ntp (systemd-timesyncd)
Turbo	PLAT_canTurbo()=false, no-ops (tg5040's turbo rides trimui_inputd; no H700 equivalent wired)
Deep sleep	PLAT_supportsDeepSleep()=1; suspend script wraps echo mem (06)
Sample rate	tg5040 logic; consults PLAT_bluetoothConnected() only — calling GetAudioSink() here segfaulted minarch pre-InitSettings() (shm not yet mapped)
WiFi/BT	generic_wifi.c / generic_bt.c + init scripts (07)

libmsettings (workspace/h700/libmsettings/msettings.c, ~1600 lines)

• Brightness: /dev/disp ioctl DISP_LCD_SET_BRIGHTNESS (0x102), args {0, raw, 0, 0}; level curve 0→4, 1→6, 2→10, 3→16, 4→32, 5→48, 6→64, 7→96, 8→128, 9→192, 10→255. Stock brightCtrl.bin is killed at launch so it can't fight us.
• Volume: tinyalsa on card 0. digital volume is a 0–63 attenuator — the code writes mixer_ctl_set_percent(digital, 0, 100 - val) (reversed mapping, confirmed correct by ear; the control's TLV metadata is garbage). lineout volume secondary. Tested on RG40XXV and RG34XXSP: UI and audible levels are correct from mute through 100%, with no current volume-control issues.
• Mute: SPK switch off + store/restore volume (no /sys/class/speaker/mute on H700). The physical volume UI still supports mute; Fn-switch settings are gated off because RG XX devices have no Fn switch.
• Color temperature: /sys/class/disp/disp/attr/color_temperature — works (tested ✅). The sibling enhance_contrast / enhance_saturation / enhance_bright attrs are also written, but have no visible effect on RG XX panels — those settings are gated off for h700 (04). All scale*() switches carry default: cases (a draft could hit uninitialized values on out-of-range input).
• DisplayCal: same gamma-LUT ioctls as tg5040 (0x10b/0x10c/0x10d) — worked 1:1 as predicted; tested passing on RG40XXV including persistence across sleep and game launch.
• HDMI: GetHDMI() probes /sys/class/extcon/hdmi/{state,cable.0/state,...} and SetHDMI() switches disp0, framebuffer/layer geometry, and the ALSA route. Menu, game, audio, and in-game hotplug in both directions pass on RG40XXV (04).
• Jack detection: not wired (path exists but historically dead — 00).
• Debug printfs were stripped/commented from the runtime path; remaining prints are error paths.

keymon (workspace/h700/keymon/keymon.c)

tg5040 structure + H700 codes: reads event1 (buttons) + event2 (volume 115/114) + event0 (power 116). MENU+vol = brightness, SELECT+vol = colortemp, plain vol = volume — tg5040 parity. No mute DIP switch on H700 (gpio243 watcher dropped). Tracks MENU on code 312 only — the synthetic 354 post-tap pulse is ignored (see MENU quirk above).

Cores (workspace/h700/cores/)

tg5040's list verbatim — the full 28-core set builds and ships (same aarch64/a53 target), with the picodrive LTO/patch-hygiene fixes noted in 01.

Shared-code integration surface

Most H700 code is platform-local, but these shared areas knowingly carry H700 integration. Review them when rebasing or adding another platform:

• workspace/all/common/generic_video.c fails fast on SDL/window/GL initialization errors instead of continuing into a black screen.
• workspace/all/common/api.c / api.h expose PWR_requestSleep() for lid-driven sleep and close audio with SND_quit() before suspend.
• workspace/all/audiomon/audiomon.cpp checks BlueALSA availability; its Makefile gives H700 the compile-time omission for the unsupported delay 0 option.
• workspace/all/minarch/makefile includes H700 in the tg5040-class RA/CHD/SRM/ sample-rate feature filters.
• workspace/all/settings/settings.cpp, btmenu.*, and the settings Makefile carry the Anbernic vendor/model capabilities and stock-Bluetooth-audio integration.
• workspace/all/syncsettings/syncsettings.c restores the display settings and gamma LUT after suspend.
• workspace/all/bootlogo/bootlogo.c contains the diagnostic empty-state handling and the optional BOOTLOGO_PREVIEW_ROTATE_CW hook used by RG28XX.
• Root/workspace makefiles register the H700 platform, dedicated h700-toolchain, rfkill, LedControl, and the H700 Bootlogo build.

Any change in this surface requires the H700 build plus the tg5040 regression build recorded as BUILD-01 and BUILD-02 in 08. A new shared button or PLAT_* hook must be defined on every platform; a full build is required because single-pak builds omit some translation units.

What disappeared vs tg5040 (as planned)

• btmanager/ (BlueZ-upgrade pakz — H700 uses firmware-provided BlueZ)
• poweroff_next/, reboot_next (systemd poweroff/reboot work fine) Reversed. That reasoning was wrong twice over: the tg5040 limbo is a PMIC latch that no init system can fix, and the base OS runs BusyBox init with no systemd at all. Both tools are now ported — see 06.
• LED animation wiring (led_anim), ledcontrol.elf (gated to tg50x0 in workspace/makefile) Partly reversed. led_anim really is TrimUI-only, but the premise that RG XX has no RGB LEDs was wrong — RG40XX H/V and RG CubeXX drive theirs from an MCU over UART5 (00). ledcontrol.elf now builds and ships for h700 as well, and LedControl is platform-aware: it asks the platform for the light count, names and effect list, and shows a "no RGB lights" screen on the models without them. Bootlogo was initially gated off too, but has since been ported: Bootlogo.pak builds for h700 with resolution-keyed preset folders (640x480, 720x480, 480x640, 720x720) selected via $DEVICE, writes bootlogo.bmp to mmcblk0p2 (BOOTLOGO_PARTITION in platform.h), backs up the stock logo as original.bmp on first apply (see 02 for the TF1-write-policy exception), and on the RG28XX rotates previews to boot orientation (BOOTLOGO_PREVIEW_ROTATE_CW, 04)
• trimui_inputd interactions (turbo, gpio 5V export)

New vs tg5040: rfkill/ — a minimal from-scratch /dev/rfkill ioctl tool (~120 lines), since the stock binary can't be relied on.