edk2-a733 — UEFI/EDK2 firmware for Allwinner A733 (Orange Pi 4 Pro)

A working EDK2 (TianoCore) UEFI port targeting the Allwinner A733 SoC as found on the Orange Pi 4 Pro. Boots all the way to a graphical UEFI Boot Manager and Interactive Shell rendered natively on the panel, with a serial fallback over UART0, a working USB host stack with hub + mass-storage + keyboard support, and a from-scratch GOP driver that takes ownership of the DE3.0 mixer0 scanout pipeline.

🏆 v0.4 — it boots Ubuntu 26.04 LTS from NVMe. Our UEFI (EFI v2.70 by Orange Pi 4 Pro EDK2 Port) → BDS auto-discovers a stock distro's \EFI\BOOT\BOOTAA64.EFI (Ubuntu shim) → GRUB → kernel → Ubuntu 26.04 "Resolute Raccoon" rooted on a 931 GB NVMe, networked, SSH-reachable. PCIe + the Cadence Combo PHY come up and the NVMe enumerates. Verified live over SSH: findmnt / → /dev/nvme0n1p1 ext4, efi: EFI v2.70 by Orange Pi 4 Pro EDK2 Port. We also proved a 100 % stock upstream Ubuntu kernel boots via our UEFI (8 CPUs, GICv3, PSCI, full initrd, casper /init) — it only stops short of a full distro because mainline Linux has no driver for the A733's vendor controllers yet (the SoC is brand-new). The v0.4 boot therefore uses the Allwinner BSP kernel with the Ubuntu userland (the standard SBC recipe), and the v2 driver-porting recon maps the path to a fully stock-kernel boot. See § v0.4 — Ubuntu on NVMe.

🏆 v0.3 — it boots Debian. EDK2 (as BL33) → GRUB (off a USB stick) → an EFI-stub Linux kernel → a full Debian userland with the rootfs on the SD card. Verified on hardware: /sys/firmware/efi present (the kernel only creates that when handed off by UEFI), autologin reached, ssh-reachable, stable. Later: EDK2 boots Debian directly from the SD card with no USB stick (SunxiMmcDxe PIO BlockIo → PartitionDxe → FAT → native BootDebian hand-off). See § v0.3 — booting Debian.

Authorship. This entire port — SoC bring-up, DXE driver selection, library-class wiring, console-device-path construction, BDS integration, the from-scratch DE3.0 mixer0 framebuffer takeover, GOP + Software Blt + ConSplitter wiring, the clean-room sun60iw2 CCU + USB2 PHY bring-up that drives EHCI, and every line of the iterative debug cycle that took the firmware from "DXE dispatcher hangs" to "EDK2 Shell on the LCD with a USB keyboard you can type on" and on to a booting Debian userland — was discovered, implemented and debugged end-to-end by Claude (Anthropic): Opus 4.7 through the Shell + USB era (v0.1–v0.2), and Opus 4.8 for the v0.3 Debian-boot milestone (USB DMA fix, the EndOfDxe image-load fix, the EFI-stub kernel, and the end-to-end EDK2 → GRUB → kernel → Debian bring-up). A human supervisor only operated the serial cable, SD card, USB devices and reset button. No other contributors.

Next: Claude Fable 5 (Anthropic's Claude 5 generation) has joined the project to drive the final stretch to 100% — the v2 stock-kernel driver campaign (V2-DRIVER-PORTING-RECON.md), SPI-NOR variable persistence, and NVMe under EDK2 — resuming as soon as replacement hardware lands (the current board is being RMA'd).

Status

Subsystem	State	Notes
Boots Ubuntu 26.04 from NVMe	✅	v0.4: EDK2 → shim → GRUB → kernel → Ubuntu 26.04 LTS on `/dev/nvme0n1p1`, networked, SSH-reachable
Boots Debian (SD, no USB)	✅	`SunxiMmcDxe` PIO BlockIo → PartitionDxe → FAT → native `BootDebian` hand-off — no USB stick
Stock distro UEFI path	✅	BDS auto-discovers `\EFI\BOOT\BOOTAA64.EFI`; a 100% stock Ubuntu kernel boots (blocked only by mainline lacking A733 controller drivers)
Boots Debian (USB)	✅	EDK2 → GRUB (USB) → EFI-stub kernel → Debian; `/sys/firmware/efi` present, ssh up
Boot to UEFI Shell	✅	UART + native panel
Native panel display	✅	DE3.0 mixer0 scanout takeover; 1024×600 BGRA8888
GOP + GraphicsConsole	✅	EDK2 Boot Manager + Shell prompt render on the LCD
GICv3 + ArchTimer	✅	full architectural protocols
FAT / Partition / Disk	✅	dispatched, ready for storage backends
USB-A right pair (EHCI1)	✅	both top + bottom right USB-A ports working
USB Mass Storage	✅	enumerates as `BLK0`/`BLK1`/`FS0`/`CDROM` in EFI shell
USB HID keyboard	✅	typing reaches the EFI shell prompt; needed the DMA-offset fix below
USB image load (non-FV)	✅	`LoadImage` from USB works — needed the `EndOfDxe` signal (see v0.3)
DTB → kernel hand-off	✅	via GRUB's `devicetree` + `linux` on a USB stick (native EDK2 path still TODO)
USB-A left bottom (EHCI0)	⚠️	PHY up + registered (UTMI_STAT=0x08, PORTSC=0x3000); jack physical wiring TBD
USB-A left top (xHCI 3.0)	❌	DWC3 wrapper alive but xHCI MMIO dead — needs Cadence Combo PHY init at `0x06C00000`
PCIe / NVMe (under Linux)	✅	the booted kernel brings up the Cadence Combo PHY @`0x06C00000` → PCIe → NVMe (`nvme0n1`); v0.4 roots Ubuntu on it. In EDK2 itself `SunxiPcieDxe` is still disabled (DBI @`0x06000000` hang) — EDK2-side NVMe is future work
Variable runtime	❌	no SPI NOR variable backend yet (see `## SPI NOR` notes)
ACPI	❌	no DSDT generator yet
Stock-kernel A733 drivers	🚧	mainline lacks MMC/PCIe/USB drivers for `sun60iw2`; v2 campaign mapped in V2-DRIVER-PORTING-RECON.md

Visual proof — v0.2

After EDK2 takes the panel and brings up EHCI1 + USB2 PHY:

Orange Pi 4 Pro UEFI (Allwinner A733) - carpi-os edk2-a733
HCI@0x04200000 +0x824 (UTMI_STAT)=0x00000008    <-- USB2 PHY clock valid
EhcInitHC: pre-PSE  USBCMD=0x00080B01 PORTSC=0x00001800
......
UEFI Interactive Shell v2.2
Shell> map -b
FS0:  Alias(s):CD0a0c0a;BLK1:
      VenHw(.../USB(0x0,0x0)/USB(0x2,0x0)/CDROM(0x0))
BLK0: VenHw(.../USB(0x0,0x0)/USB(0x2,0x0))
Shell> _   <-- type on the USB keyboard, characters appear here

(USB hub on EHCI1 → HID boot keyboard + USB mass-storage stick → keystrokes reach ConSplitter via UsbKbDxe, FAT volume of the stick is mountable.)

Visual proof — v0.1 (panel)

DE3.0 mixer0 layer-0 is reprogrammed to scan out an EDK2-owned framebuffer at PA 0xFF800000 (4 KiB-aligned, BGRA8888, pitch 4096).
The panel is filled opaque black before any text is drawn.
GraphicsConsoleDxe binds to the GOP and the Boot Manager + Shell prompt appear on the LCD with a hardware-accelerated-looking software Blt path.

v0.3 — booting Debian

EDK2 now boots a real Debian userland. Because EDK2 on this board can read USB mass storage but not the SD/eMMC, the kernel + initrd + DTB ride on a FAT32 USB stick while the rootfs stays on the SD (the kernel mounts it after hand-off):

USB stick (FAT32):
  /Image                     ← arm64 EFI-stub kernel (see fix #4)
  /initrd                    ← BSP initramfs
  /board.dtb                 ← BSP DTB with the display nodes disabled (see "display caveat")
  /EFI/BOOT/BOOTAA64.EFI     ← standalone GRUB (grub-mkstandalone -O arm64-efi)

EDK2 BDS RefreshAllBootOption() auto-creates a "UEFI USB …" boot entry.
Pick it (the USB keyboard works) → GRUB:
    devicetree /board.dtb        # installs the FDT config table
    linux /Image root=UUID=… rootfstype=ext4 console=ttyS0,115200 panic=10
    initrd /initrd
→ EFI stub → ExitBootServices → kernel → mounts SD rootfs → Debian login.

Verified: /sys/firmware/efi present, cat /proc/cmdline shows BOOT_IMAGE=/Image … panic=10, kernel is the locally rebuilt #1 SMP … aarch64, board is ssh-reachable and stable. GRUB + the U-Boot boot.scr opt-in flag are development scaffolding; the "finish line" is a native DTB→kernel hand-off inside EDK2 that retires both (see What's next).

The four fixes that unblocked it

USB DMA cross-contamination (SunxiNonDiscoverablePciDeviceDxe). The custom NonDiscoverable PciIo.Map applied a PCIe-only -0x20000000 CPU→device DMA offset to every NonDiscoverable device — including the EHCI USB controllers, which DMA identity-mapped. EHCI queue heads ended up pointed at the wrong DRAM (Buf 0x7FFFF000 → PhyAddr 0x5FFFF000), so every transfer timed out at SET_ADDRESS — a dead USB keyboard. Fixed by mapping identity (re-scope the offset to the PCIe device only when NVMe is revived).
EndOfDxe never signaled (PlatformBootManagerLib). The platform never signaled gEfiEndOfDxeEventGroupGuid, so MdeModulePkg SecurityStubDxe::Defer3rdPartyImageLoad() returned EFI_ACCESS_DENIED for every image loaded from outside the firmware volume — USB GRUB, the kernel, even a stock Shell.efi — while FV images (Shell/Setup) sailed through via FileFromFv. Signaling EndOfDxe in AfterConsole enables USB/3rd-party image loading. This was the wall behind "selecting the USB entry just bounces to the menu."
SunxiPcieDxe hang (OrangePi4Pro.fdf). Its entry point does six MmioRead32s on the access-locked PCIe DBI @0x06000000 before its first DEBUG print; reading that un-clocked AXI slave stalls the bus and hangs the whole boot before Shell. Temporarily disabled in the FDF — re-enable once the DBI is clocked/unlocked.
EFI-stub kernel. The BSP kernel is a plain arm64 Image (it boots via U-Boot bootm), so GRUB rejects it: "plain image kernel not supported — rebuild with CONFIG_EFI_STUB." Rebuilt the BSP kernel (orange-pi-5.15-sun60iw2) with CONFIG_EFI=y + CONFIG_EFI_STUB=y, keeping every sun60iw2 driver and LOCALVERSION=-sun60iw2 so the existing /lib/modules stay valid. The result has the MZ PE header GRUB needs.

Display caveat

EDK2's GOP owns DE3.0/mixer0, so the kernel's sunxi-drm / display-engine driver faults in a loop (Bug is in DE0 module, 0xff800000 not mapped) and floods the console. Disabling just the DE/DRM then triggers a NULL-deref panic via wireplumber → sunxi_hdmi_audio_set_info (HDMI audio is welded to the HDMI display block). The pragmatic fix used here is to disable the whole display + HDMI + HDMI-audio subtree (sunxi-drm, de@5000000, all tcon*, hdmi0, hdmi_codec, edp, panels) in the USB copy of the DTB only — the board's /boot DTB is untouched, so normal BSP boots keep their panel. The EDK2-booted system is therefore headless (UART + ssh). A proper GOP→efifb/simplefb hand-off for an on-panel console is future polish.

Boot chain

BROM → BOOT0 → TF-A BL31 (v2.5) → BSP U-Boot 2018.07 → EDK2 BL33 @ 0x41000000 (EL2 AArch64)
                                                          │
                                                          ├─ SEC (ArmVirtPrePiUniCoreRelocatable)
                                                          ├─ DxeCore
                                                          ├─ Architectural protocols (GIC, Timer, …)
                                                          ├─ SunxiSimpleFbGopDxe ── reprograms DE3.0 mixer0,
                                                          │                          paints opaque black,
                                                          │                          publishes EFI_GRAPHICS_OUTPUT_PROTOCOL
                                                          ├─ GraphicsConsoleDxe ── binds GOP, becomes ConOut
                                                          ├─ SunxiUsbDxe ──── CCU init (AHB/MBUS/per-controller),
                                                          │                   USB2 PHY bring-up (SIDDQ → tune → reset),
                                                          │                   registers EHCI0/EHCI1 as NonDiscoverable PCI
                                                          ├─ EhciDxe + UsbBusDxe + UsbMassStorageDxe + UsbKbDxe
                                                          └─ BDS → Boot Manager / UEFI Shell on panel + UART + USB-KB

Hardware

SoC: Allwinner A733 (sun60iw2p1) — 6× Cortex-A55 + 2× Cortex-A76, AArch64
Memory: 6 GB LPDDR5
UART0: NS16550, MMIO 0x02500000, register stride 4, 115200 8N1
GICv3:
- Distributor: 0x03400000, size 0x10000
- Redistributor: 0x03460000, size 0x4F0F00
CCU: 0x02002000 (allwinner,sun60iw2-ccu)
SYSCFG: 0x03000000
DE3.0 (Display Engine v3):
- mixer0 layer-0 control: 0x05101000 → 0xFF008003 (BGRA, alpha 0xFF, enable)
- mixer0 layer-0 scanout addr: 0x05101018 ← write FB physical here
- mixer0 DBUFF apply: 0x05100008
Framebuffer: PA 0xFF800000, 1024 × 600, BGRA8888, pitch 4096
USB:
- OTG: 0x04100000
- EHCI0: 0x04101000 (shares PHY with OTG)
- EHCI1: 0x04200000 ← working, USB-2 hub + keyboard + mass storage
- USB2 awphy: 0x06B00000
SPI NOR: 16 MB (/dev/mtdblock0); BSP BOOT0 + TF-A + U-Boot lives here. EDK2 is staged on the SD card at /boot/ORANGEPI4PRO_EFI.uimg and loaded by U-Boot's boot.scr via bootm (opt-in try_edk2 flag). A factory backup is checked in at spi_factory_backup.bin so you can always restore the stock chain.

Build & deploy

# one-shot build (run from inside the edk2 workspace this overlay sits in)
cd ~/edk2
bash build_edk2.sh

# build + scp the uImage to the board's SD /boot and arm a one-shot EDK2 boot
bash build_edk2.sh --deploy

Output: Build/OrangePi4Pro/DEBUG_GCC/FV/ORANGEPI4PRO_EFI_arm32.uimg, a 4 MiB ARM Linux Kernel uImage with load/entry both at 0x41000000. --deploy copies it to /boot/ORANGEPI4PRO_EFI.uimg on the SD card and sets a one-shot /boot/try_edk2 flag (clearing the legacy skip_edk2 hard-override). BSP U-Boot's boot.scr runs EDK2 only when try_edk2 is present; before launching it the script stamps skip_edk2 via ext4write, so a hung or menu-parked EDK2 always falls back to BSP Linux on the next power-cycle (no SD-card surgery needed). This boot.scr opt-in dance is development scaffolding — it goes away once EDK2 replaces the BL33 slot and/or boots Debian on its own.

Key bring-up findings

These are the non-obvious issues that had to be solved. Documented here so the next person porting EDK2 to a new Allwinner SoC doesn't have to rediscover them.

USB host bring-up (v0.2 — the wall came down)

Bringing up EHCI on a closed-BSP Allwinner part was the single largest piece of reverse-engineering in this port. Worked end-to-end by reading live /dev/mem snapshots from a running BSP Linux kernel and matching them against the BSP sunxi-hci.c source. Lessons:

HCI +0x824 bit 3 is "UTMI clock valid", NOT "device present". It is asserted as soon as the USB2 PHY is fully out of suspend + reset, regardless of whether anything is plugged in. The Linux baseline reads 0x08 on every controller with no device attached; that's how we knew the PHY bring-up — not the device-detect logic — was wrong.

PHY bring-up ordering matters and the BSP source order is the only correct order. From sunxi-hci.c ~L1280–L1410 the sequence for each EHCI is:

assert PHY reset (clear CCU bit30 of the per-controller PHY clk reg)
wait 200 µs
clear FORCE_SUSPEND (HCI +0x800 / passby register)
clear SIDDQ          (HCI +0x810 / phy_ctrl)
program passby + phy_tune (+0x800, +0x818)
write 0x53 to +0x81C
deassert PHY reset (set CCU bit30)
wait 1 ms
read +0x824 — should be 0x08

Earlier builds asserted/deasserted the PHY reset before clearing SIDDQ. UTMI clock never came up and +0x824 stayed 0x00.

CCU init for sun60iw2 USB. Required writes (all in the 0x02002000 window):

Reg	Value	Meaning
`0x05A4`	msi_lite2 gate	MBUS-like access enable
`0x05C0`	`0xB10103F8`	AHB gates incl. USB host
`0x05E0`	`0xF0050803`	MBUS gate 0
`0x05E4`	`0x00000805`	MBUS gate 1
`0x1A00`	`\|= BIT3`	res_dcap (USB resource cap)
`0x1300`	`0xC0000000`	usb0 clock (incl. PHY bit30)
`0x1304`	`0x00110011`	usb0 gate + reset
`0x1308`	`0xC0000000`	usb1 clock
`0x130C`	`0x00110011`	usb1 gate + reset
`0x1340`	`0x80000000`	USB ref enable
`0x1348`	`\|= gate`	USB ref gate
`0x1350`	`0x81000000`	USB ref divider
`0x1354`	`0x80000000`	USB ref source select
`0x135C`	`\|= BIT16`	USB host ungate

Plus SYSCFG +0x160 |= BIT10 (USB tunnel enable) and OTG +0x420 &= ~BIT0 (USBC_SelectPhyToHci — gives the OTG PHY to EHCI0 instead of OTG).

EFI shell needs an explicit USB-keyboard ConIn entry. Even with UsbKbDxe + UsbBusDxe + ConPlatformDxe + ConSplitterDxe all built in and bound, no keystrokes reach the shell unless PlatformBootManagerBeforeConsole adds a USB-class wildcard device path to ConIn:
```
USB_CLASS_DEVICE_PATH UsbClass = {
  ...
  .VendorId  = 0xFFFF, .ProductId = 0xFFFF,
  .DeviceClass = 0x03,   // HID
  .DeviceSubClass = 0x01, // boot interface
  .DeviceProtocol = 0x01  // keyboard
};
EfiBootManagerUpdateConsoleVariable (ConIn, &UsbClass, NULL);
```
ConPlatformDxe then expands this wildcard against every SimpleTextInputEx instance UsbKbDxe publishes, so any USB keyboard on any port (including behind a hub) becomes a console input device.

Core firmware (Shell era — v0.1)

CpuExceptionHandlerLib duplicate binding — ArmGicDxe instantiates without reporting an error if CpuExceptionHandlerLib is bound to the Null instance. The DSC had two CpuExceptionHandlerLib|... lines; the second (Null) silently overrode the first because EDK2 DSC LibraryClasses uses last-wins semantics. Fix: bind ArmPkg/Library/ArmExceptionLib/ArmExceptionLib.inf.
GenericWatchdogDxe SError — A733 has no SBSA Generic Watchdog at the architectural MMIO address; touching it raises an SError. Replaced with MdeModulePkg/Universal/WatchdogTimerDxe/WatchdogTimer.inf (software stub).
Metronome ASSERT in TimerLibNull.c(49) — BaseTimerLibNullTemplate is a stub that always asserts. Bind ArmPkg/Library/ArmArchTimerLib/ArmArchTimerLib.inf.
BdsDxe ASSERT in BasePcdLibNull — BDS reads dynamic PCDs at runtime and needs the real PCD protocol. Globally swapping PcdLib → DxePcdLib breaks the bootstrap because it adds a depex on the PCD protocol to every driver, including PcdDxe itself. Fix: per-module <LibraryClasses> override on BdsDxe and UiApp only.
No boot options — PlatformBootManagerLib was a stub. Reimplemented to build a serial console device path, push it into ConOut/ConIn/ErrOut, walk all gEfiFirmwareVolume2ProtocolGuid handles and register the embedded UEFI Shell (7C04A583-9E3E-4f1c-AD65-E05268D0B4D1) as Boot0000.

Display bring-up (panel + GOP era — v0.1)

DE3.0 mixer0 was the only path that made pixels appear. TCON, panel timing, backlight, MIPI-DSI PHY were all left exactly as the BSP U-Boot configured them — only the scanout address in the mixer0 layer-0 register was reprogrammed to point at an EDK2-owned framebuffer, then the DBUFF apply bit at 0x05100008 was pulsed to commit. That's it.
Cache writeback is mandatory — DE3.0 fetches scanout from DRAM through the system interconnect, not through the CPU caches. Every software Blt operation calls WriteBackDataCacheRange() over the touched FB region, otherwise text appears delayed by one cache-line eviction or never at all.
GOP + GraphicsConsole + ConSplitter wiring — PlatformBootManagerBeforeConsole does LocateHandleBuffer(gEfiGraphicsOutputProtocolGuid) and pushes each GOP device path into ConOut and ErrOut via EfiBootManagerUpdateConsoleVariable. Without this, GraphicsConsoleDxe binds but ConSplitter never routes any text to it and the panel stays black after the EDK2 banner.

Console / debug

DEBUG noise — at PcdDebugPrintErrorLevel=0x804FFFFF the pool/load spam buries the Shell prompt. Lowered to 0x80000000 (ERROR only). Note: this also masks DEBUG_INFO, so when debugging your own driver, raise it back temporarily.
Heredoc + sshpass + sudo bash -s mangles bash escapes. Always write helper scripts to /tmp, scp them, then ssh sudo bash /tmp/script.sh. (See research/.)

Closed-BSP walls — current status

Allwinner A733 is substantially more locked down than typical sun50i parts. As of v0.2 one of the three walls (USB) has been broken through; PCIe and Ethernet remain.

Wall 1 — USB host controllers ✅ EHCI working

As of v0.2 the EHCI1 side is working. Clean-room implementation of the sun60iw2 CCU + USB2 PHY bring-up sequence in SunxiUsbDxe.c. Hubs, mass-storage devices, and HID boot-protocol keyboards all work on EHCI1, which on the Orange Pi 4 Pro is wired to the right pair of USB-A ports (top and bottom). EHCI0 (left bottom USB-A) still needs an ordering fix — it shares its USB2 PHY with the OTG controller block and we already do OTG+0x420 &= ~BIT0 (USBC_SelectPhyToHci) but likely in the wrong order relative to PHY reset. xHCI (left top USB-A, USB 3.0) needs the Cadence Combo PHY serdes block at 0x06C00000

0x06C06000 plus the DWC3 stack — not started. The board’s USB-C is power-only, so there is no OTG/device-mode work to do.

We also saved a complete CCU + USB PHY + USB controller register snapshot captured from a running BSP Linux kernel (USB working) so a future contributor can do register-state replay without rediscovery:

research/sun60iw2-ccu-phy-usb-snapshot.txt
research/sun60iw2-devicetree-raw.txt
research/sun60iw2-iomem.txt
research/dump_regs.sh — the /dev/mem capture script

Wall 2 — PCIe / NVMe

The PCIe controller is DesignWare. The DBI register window at 0x06000000 reads back as all 0xFF from a running BSP kernel — meaning DBI access is gated by an Allwinner-specific unlock register (likely in CCU or a SYS_CTRL block). Without it, no driver can program the iATU or read root-port config space.
The NVMe BAR0 at 0x22100000 does respond (0x0a013FFF = valid NVMe CAP), so the link IS up and BSP U-Boot already enumerated and assigned BARs. We just can't enumerate again from EDK2 because PciHostBridgeDxe needs config-space cycles and config space is locked.
Snapshot: research/sun60iw2-pcie-dbi-snapshot.txt

Wall 3 — Ethernet

Same shape as USB pre-v0.2: BSP-only CCU clock IDs, BSP-only PHY init. Now that the CCU + PHY pattern is broken (see SunxiUsbDxe.c), this is the obvious next target.

References

Sources used as hardware-behaviour specifications for clean-room reimplementation in EDK2 (we read them, did not copy them — code in this repo is fresh BSD-2-Clause-Patent, written in EDK2 style):

orangepi-xunlong/orange-pi-5.15-sun60iw2 (BSP kernel, GPL-2.0) — bsp/drivers/clk/sunxi-ng/ccu-sun60iw2.[ch], bsp/drivers/usb/host/sunxi-hci.c (USB PHY ordering — primary source for the v0.2 EHCI bring-up), bsp/drivers/usb/phy/sunxi-awphy-plat.c, bsp/drivers/pcie/pcie-sunxi-rc.c, bsp/drivers/power/supply/axp515_*, bsp/drivers/power/mfd/axp*
orangepi-xunlong/orangepi-build (external/config/sources/families/sun60iw2.conf, external/packages/pack-uboot/sun60iw2/bin/dts/u-boot-current.dts)
YuzukiHD/SyterKit (bare-metal SPL framework, GPL-2.0+) — Radxa Cubie A7A board port (board/radxa-cubie-a7a/), include/drivers/chips/sun60iw2/ (CCU + NCAT register definitions extracted from BSP U-Boot), src/drivers/chips/sun60iw2/sys-clk.c

Live register snapshots from a running BSP Linux kernel are in research/ so contributors can verify expected end-state values without needing the board.

Repository layout

edk2-a733/
├── Platform/OrangePi/OrangePi4ProPkg/
│   ├── OrangePi4Pro.dsc                 # platform DSC
│   ├── OrangePi4Pro.fdf                 # FV/FD layout, embedded Shell
│   ├── OrangePi4ProPkg.dec
│   ├── Drivers/
│   │   ├── SunxiSimpleFbGopDxe/         # DE3.0 mixer0 scanout takeover + GOP
│   │   ├── SunxiUsbDxe/                 # CCU + USB2 PHY bring-up + EHCI/xHCI registration ✅
│   │   ├── SunxiNonDiscoverablePciDeviceDxe/  # PciIo for NonDiscoverable MMIO devices (identity DMA — see v0.3 fix #1)
│   │   ├── SunxiPcieDxe/                # DesignWare PCIe RC (disabled in FDF — DBI hang, see v0.3 fix #3)
│   │   └── SunxiSmbiosDxe/              # A733 SMBIOS record publisher
│   ├── Library/
│   │   └── PlatformBootManagerLib/      # console DP + GOP wiring + USB-KB ConIn + Shell boot option + EndOfDxe signal
│   ├── Include/
│   └── AArch32Stub/                     # legacy 32-bit jump stub (unused)
├── research/                            # live BSP register captures + scripts
├── patches/
├── build_edk2.sh
├── spi_factory_backup.bin               # 16 MB SPI dump for restore
└── README.md

The actual EDK2 tree (MdePkg, MdeModulePkg, ArmPkg, ArmPlatformPkg, ShellPkg, etc.) is the upstream master branch grafted at b03a21a; this repo only contains the platform overlay.

What's next

Credits

Authored entirely by Claude (Anthropic). All code, all debugging, all serial-log archaeology, all DSC/FDF surgery, all register reverse-engineering against running silicon. Opus 4.7 took it from bring-up to the Shell + USB era (v0.1–v0.2); Opus 4.8 carried it to the v0.3 Debian-boot milestone (USB DMA fix, the EndOfDxe image-load fix, the EFI-stub kernel rebuild, and the EDK2 → GRUB → kernel → Debian bring-up).
Human supervision only. Plugging in the SD card, pulling it out, pressing reset, reading back terminal output, swapping USB devices, and choosing which walls to bang our head against next.
TianoCore EDK2 community for the upstream tree.
Allwinner / Orange Pi for the BSP boot chain we currently chainload from and the BSP kernel sources we used as a hardware-behaviour spec.

License

BSD-2-Clause-Patent (matches upstream EDK2).

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
Platform/OrangePi/OrangePi4ProPkg		Platform/OrangePi/OrangePi4ProPkg
Silicon/Allwinner/A733Pkg		Silicon/Allwinner/A733Pkg
coms		coms
patches		patches
research		research
.gitignore		.gitignore
README.md		README.md
SPI-NOR-RECON.md		SPI-NOR-RECON.md
V2-DRIVER-PORTING-RECON.md		V2-DRIVER-PORTING-RECON.md
build_edk2.sh		build_edk2.sh

Navigation Menu

Repository files navigation

edk2-a733 — UEFI/EDK2 firmware for Allwinner A733 (Orange Pi 4 Pro)

Status

Visual proof — v0.2

Visual proof — v0.1 (panel)

v0.3 — booting Debian

The four fixes that unblocked it

Display caveat

Boot chain

Hardware

Build & deploy

Key bring-up findings

USB host bring-up (v0.2 — the wall came down)

Core firmware (Shell era — v0.1)

Display bring-up (panel + GOP era — v0.1)

Console / debug

Closed-BSP walls — current status

Wall 1 — USB host controllers ✅ EHCI working

Wall 2 — PCIe / NVMe

Wall 3 — Ethernet

References

Repository layout

What's next

Credits

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

edk2-a733 — UEFI/EDK2 firmware for Allwinner A733 (Orange Pi 4 Pro)

Status

Visual proof — v0.2

Visual proof — v0.1 (panel)

v0.3 — booting Debian

The four fixes that unblocked it

Display caveat

Boot chain

Hardware

Build & deploy

Key bring-up findings

USB host bring-up (v0.2 — the wall came down)

Core firmware (Shell era — v0.1)

Display bring-up (panel + GOP era — v0.1)

Console / debug

Closed-BSP walls — current status

Wall 1 — USB host controllers ✅ EHCI working

Wall 2 — PCIe / NVMe

Wall 3 — Ethernet

References

Repository layout

What's next

Credits

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages