Experimental port of macOS 26.5 Tahoe (25F71) to the Google Pixel 8a ("akita": Google Tensor G3 "zuma" SoC,
Mali-G715 MC7 GPU). The full OS boots from RAM (fastboot boot, nothing is flashed) to the desktop.
WindowServer composites on the Mali GPU through Metal → metal2vulkan → Mesa PanVK, driven by our own IOKit
Mali CSF driver in the kernel. The touchscreen works, with gestures.
⚠️ Research project, not a daily driver. This repository contains no Apple binaries and no ARM firmware: get them legally yourself (Apple's IPSW, the vendor partition of your own phone).
| Milestone | Status |
|---|---|
XNU (DEVELOPMENT kernel) boots on the Tensor G3, from RAM with fastboot boot |
✅ |
| macOS Recovery with menu bar, Disk Utility, Terminal, Safari | ✅ |
| Full macOS to the desktop (Finder, Dock, System Settings, Terminal), autologin | ✅ |
| Own IOKit driver for the Mali-G715 (CSF: MCU firmware, MMU, queue groups) | ✅ |
| Vulkan (Mesa PanVK ported to Darwin arm64e) | ✅ |
Metal over Vulkan (metal2vulkan): MTLCreateSystemDefaultDevice = "Mali-G715 MC7" |
✅ |
| WindowServer composites on the GPU (SkyLight's Metal compositor) | ✅ |
| Dark mode, Liquid Glass (transparency) and wallpaper at boot | ✅ |
| Touchscreen (Goodix on SPI, our own driver) with gestures: tap, drag, long press = right click, two-finger scroll | ✅ |
| Window dragging: 15.7 → ~22 → ~50–55 fps (batched GPU submits, early swap completion, PanVK fast paths; 10-minute drag test: median 55.6 fps) | ✅ (goal: 60) |
| GPU stable under stress (ring-full fault fixed; two jobs of slack counted from the job the GPU is in) | ✅ |
| Real clock: boots with the time from the PMIC's RTC (S2MPG14), kept in sync | ✅ |
| Autologin straight to the desktop with the persistent session | ✅ |
| CPU thermal regulator (step down at 85 °C, step back up below 75 °C) | ✅ |
Persistent home and settings: the user's files survive a reboot (APFS volume in a window of userdata) |
✅ |
| Swap on flash (real swapfiles, 2–6 GB under pressure) | ✅ |
| RAM: macOS sees 5.98 GiB (was 4 GiB; UEFI heap moved, all 8 GiB declared, integrity-tested) | ✅ |
Boot: fastboot boot → desktop in ~2 min (was ~6.5): no 30 s log pages in DarwinBoot, RELEASE UEFI, rc.server 55 → 19 s (docs/boot-time.md) |
✅ |
Persistent shader cache (/private/var/folders on the persistent container): WindowServer's pipelines 33 s → 154 ms per boot |
✅ |
| Battery in the menu bar (MAX77759 fuel gauge as an IOPMPowerSource: percent, charging, voltage, temperature) | ✅ pmset -g batt shows InternalBattery-0; read-only |
| Display brightness (MIPI DCS 0x51 to the OLED panel through DSIM0) | ✅ macOS's own slider in Control Center (AkitaBacklightService + tools/akita-brightnessd); the volume keys are macOS brightness keys (consumer usages 0x6F/0x70: HUD and slider follow) |
| Disk images (.dmg: DiskImageMounter, hdiutil) | ✅ AppleDiskImages2, the DiskImages family and HFS+ linked into the kernel; /Volumes writable |
| USB host: keyboard and mouse through a USB-C hub (our xHCI driver, HID boot protocol) | ✅ plug the hub in and they work in ~20 s (the PD engine starts by itself once the Mac link is gone); mouse cursor drawn; scrolling works; low-speed devices pending; going back to the Mac's USB needs a reboot |
| Stability: hour-long soaks with Wi-Fi/USB checks, 10-minute drag test, critical memory pressure (13.8 GB swap) | ✅ (the "hangs" after 1 h were the watchdog boot-arg akita-wdt-pet-seconds=3600: now pets for good) |
| 60 fps (WindowServer's main thread is CPU-bound in the Metal bridge) | 🔧 in progress |
Wi-Fi (BCM4383 on PCIe): joins WPA2 at boot, DHCP, internet (DNS, HTTPS); kernel log, shell and fastboot reset over Wi-Fi; en1 now listed in System Settings > Network (published in IOKit as an Ethernet-type interface); no Wi-Fi menu yet |
✅ works, docs/wifi-bringup.md |
| USB-C: Power Delivery and DisplayPort alternate mode (MAX77759 TCPC) | ✅ PD contract, host role, DP mode configured with a USB-C HDMI hub |
| DisplayPort / HDMI out (monitor mode) | ✅ macOS desktop at 1920×1080 on an external monitor through a USB-C HDMI hub, link up on plug, down on unplug, back on re-plug without rebooting (opt-in: akita-dp=1 akita-dp-monitor=1), see docs/displayport.md |
More photos in photos/.
WindowServer (SkyLight, CompositorMetal) ─┐
CoreAnimation / apps ├─ Metal.framework
│ └─ NativeMaliMetal.bundle (Metal plugin, /System/Library/Extensions)
│ └─ NativeBridgeMetalFrontend (UniXNU) + metal2vulkan (Rust)
│ └─ Mesa PanVK (libvulkan_panfrost.dylib, ported to macOS arm64e)
│ └─ darwin-drm (libdrm stand-in: Panthor ioctls → IOKit)
│ └─ MaliClient (IOUserClient)
XNU kernel ───────────────────────────────┘ └─ Mali.cpp: Mali-G715 CSF driver + IOAccelerator
NativeGOPDisplay: copies frames to the panel
What WindowServer needed to use the GPU (details in the plugin code and the kernel patches):
- Metal discovery: the driver publishes an
IOAcceleratorwithMetalPluginName = NativeMaliMetal; Metal loads the bundle and callsinitWithAcceleratorPort:. - WindowServer is a platform binary with library validation and a sandbox: with the boot-arg
akita-mali-platform=1the kernel marks the ad-hoc signed bundle as platform code and lets it openMaliClient, skipping the MACF check and the sandbox filter. [MTLDevice acceleratorPort]: CoreDisplay ties each display to the accelerator this method returns. Our device returned 0, so SkyLight fell back tocompositor_sw. The plugin now answers with the realIOAcceleratorservice.WSSystemCanCompositeWithMetal: on 25F71 this SkyLight check stays "no" even with a device. As a workaround the plugin marks it passed when loaded inside WindowServer (only on 25F71, after checking the expected instructions).
- Batched queue submits (
MALI_GROUP_SUBMIT_BATCH): one IOKit call per ~32 jobs instead of one per job (each call cost ~36 µs, ~170 per frame). Cross-queue waits are done by the GPU (SYNC_WAIT64), not by the CPU. - Flush-ID page mapped read-only to user space, so PanVK skips cache flushes the GPU already did.
- Early swap completion in the display driver: the swap completes when it is taken and the copy waits on the GPU fence.
- PanVK: 128-byte alignment for tiled images (Metal's alignment), so the bridge no longer keeps shadow copies of heap textures; render target and push-constant descriptors built locally and stored once (GPU memory is uncached).
- GPU ring protection: a queue's 64 KiB ring is never filled completely (a full ring raised
CS_CONFIG_FAULT). - Profiling tools for live stacks of WindowServer:
scripts/stacks.py,allstacks.py,sysyms.py, anddrag60.pyto measure window-drag fps over the USB input channel.
The remaining limit is CPU time in the Metal → Vulkan bridge on WindowServer's main thread (~150 µs per draw).
The system volume stays read-only, as on a Mac; the user's home now lives on a persistent APFS volume and swap goes to
flash. NativeUfs publishes a 64 GiB window of userdata as a disk (AkitaDataDisk.h, boot-arg akita-data-disk=2);
rc.server formats it on first use (only if it is blank), mounts the home at /Users/akita and the swap volume at
/System/Volumes/VM. macOS now sees 5.98 GiB of the phone's 8 GiB. Details in
docs/storage-and-memory.md.
The Pixel 8a's touch controller is a Goodix "nottingham" (Berlin family) on SPI. The firmware runs from the chip's own flash,
so the driver only powers it and reads touches (it never writes the chip's flash). platform-akita/drivers/Acpm/AkitaTouch.inc:
- rails through the PMIC (S2MPG14 via ACPM): LDO25M 1.8 V and LDO4M 3.0 V; reset and interrupt GPIOs;
- USI/SPI master setup, chip info at
0x10070, touch data read and acknowledged when the interrupt line drops; - a gesture state machine: tap = click, drag, long press = right click, two fingers = scroll wheel;
- delivery into
NativeRecoveryInput's absolute HID tablet, rotated to the landscape desktop.
Protocol notes in docs/research-touch.md.
| Path | What it is |
|---|---|
patches/xnu-12377.121.6-akita.patch |
Changes on XNU 12377.121.6 (macOS-ForSheng's and akita's: full-OS rc.server, MaliClient, platform, dark mode at boot, akita-tfp…). SSH keys are placeholders: generate your own. |
patches/macOS-ForSheng-ca57db7.patch |
Changes on CrossDarwin/macOS-ForSheng ca57db7. |
platform-akita/ |
akita platform for macOS-ForSheng (goes in platform/akita/): Mali CSF driver, ACPM (CPU clocks, thermal regulator, PMIC) and touch, UFS, input, ADT, image packaging, display driver patches (patches/display-*) and the ArmCpu patch. |
patches/mesa-2648b20-akita.patch |
Mesa main 2648b20: PanVK on Darwin (kmod over MaliClient, memory, cache, fast paths). |
darwin-drm/ |
libdrm stand-in that serves PanVK's Panthor ioctls and DRM syncobjs through the Mali IOKit user client (batched submits, GPU-side waits, flush-ID page). |
patches/UniXNU-a3faa4a.patch |
Changes on UniXNU a3faa4a, including its AdrenoMetal Metal frontend and bridge (8160223) and a new metal2vulkan patch (0031). |
patches/metal2vulkan-624bab39-akita-heap-diag.patch |
A diagnostic in metal2vulkan's heap textures (why a texture gets a shadow image). |
plugin/ |
NativeMaliMetal.bundle: the Metal plugin (UniXNU frontend + metal2vulkan + PanVK), with live profilers. |
tools/ |
Tests (Vulkan, Metal, metaldiscover) and WindowServer inspection (wsvars, wswatch, wssample, wsstack, wsfind…). |
scripts/ |
Image build (build-fullos.sh), Mali/ACPM/bridge/darwin-drm/PanVK builds, test cycles, USB shell/log, profilers. |
uefi-akita/ |
The akita platform for Mu-Silicium (akitaPkg, S5P9865Pkg): memory map, UFS, keypad. patches/Mu-Silicium-213bce4-akita.patch has the changes to its common code. |
docs/ |
Practical guide, full-OS plan, research notes (GPU, touch, USB-C video), storage and memory, development log. |
The scripts contain absolute paths of the development machine (/Volumes/datos2/unixnu-akita/...): adapt them.
- Pixel 8a with an unlocked bootloader. Always booted with
fastboot boot(RAM); no flashing needed. - Apple's macOS 26.5 (25F71) IPSW for the system and the reference kernelcache.
- CSF firmware from your own phone:
adb pull /vendor/firmware/mali_csffw-r54p2.bin, thenscripts/make-mali-csffw-inc.pygeneratesplatform-akita/drivers/Mali/mali_csffw.inc(not redistributed). - macOS on Apple Silicon with Xcode to build XNU, Mesa (meson/ninja), Rust 1.98 (metal2vulkan with
-Z build-std).
- macOS-ForSheng: the base for booting macOS 26 on non-Apple hardware.
- UniXNU: native Metal frontend and metal2vulkan bridge.
- MaliMetal: darwin-drm, the Mali user-space stack for Darwin we built on.
- Mesa: PanVK / panthor.
- Google's goodix_touch driver (GPL): reference for the touch protocol.
- nullarc: reference for a GPU WindowServer over Vulkan.
- ipsw: dyld shared cache analysis.
- Apple: XNU is open source under APSL 2.0. macOS, its frameworks and its trademarks belong to Apple; this project is not affiliated with Apple or Google.
Much of the late-2026 work (GPU performance, touch, display) was done with Claude (Anthropic); those local changes to PanVK
and the bridge are marked akita (Generated-by: LLM) and are not submitted upstream.
Our own code is GPL-3.0 (see LICENSE), compatible with macOS-ForSheng. Patches on third-party projects keep each
project's license (XNU: APSL 2.0; Mesa: MIT; UniXNU, darwin-drm and the drivers in platform-akita: BSD-2-Clause-Patent, see
each folder's LICENSE). See THIRD_PARTY.md.