Apple's open-source XNU kernel on Android phones: its own loader and drivers, Metal on the Adreno. Runs from RAM; a reboot gives the phone back to Android.
C
On the phone it runs and is exact, and it ran on softpipe: there was no GPU driver
vcvkk58593541a27e
diff --git a/docs/air.md b/docs/air.mdindex 1d53265..c32fc20 100644--- a/docs/air.md+++ b/docs/air.md@@ -469,6 +469,22 @@ there. So the whole chain works on the target operating system: the container, ttranslation, Gallium, the draw and the readback. What is left is one substitution — softpipe forir3 on the Adreno — and that needs the phone.+**On the phone, with the root on the flash, it runs and is exact — and it ran on softpipe.** The+first run of it (19 October, `logs/232842-run.log`) booted the flash-root recipe properly (both+`[ ok ] mount the iOS volume read only` and `[ ok ] the iOS volume is the root, the RAM part of+the boot on /private/preboot`), translated `clearFragment01` and read back the same four exact+floats as everywhere else. But its first line said `on softpipe`, and the kernel said why:++```+adreno: THE GPU IS NOT STARTED: the small image has no fw/a660_gmu.bin+ (the phone's /vendor/firmware/a660_gmu.bin goes into fw-gpu/)+```++The GPU driver had not started at all, because this working copy had no `fw-gpu/` — the phone's+own firmware, which is not in the repository. **Not the driver refusing work: no driver.** This is+exactly what that first line is for: the pixel was right, and reporting it as a result on the+Adreno would have been false. Read the line before the numbers.+What is *not* known until it is tried is whether the kernel's GPU driver and the submit pathbehave for a draw that does not come from CoreAnimation: the GLES path is that driver's onlyconsumer today. `rc.jobs.air` asks with the GPU idle; `rc.jobs.air2` asks again after SpringBoard@@ -533,6 +549,11 @@ drivers and the port's own Mesa are built on. The durable way to get the `ir3` nupgrade is `air_run` against a freedreno screen, which means on the phone, because freedrenoneeds a real `msm` device and this machine has none.+A working copy also needs the phone's own firmware for the GPU to start at all: `fw-gpu/`+(`a660_gmu.bin`, `a660_sqe.fw`, the `a660_zap.*` pieces), which is not in the repository and which+`tools/new-workspace.sh` does not copy. Without it the kernel says so plainly and the system falls+back to software, so a run that looks fine can be a run on no GPU.+Needed on this machine: meson, ninja, bison, flex, python3 with venv, and LLVM's `llvm-as` /`llvm-dis` for the reader check (22.1.8 here). The host build also patches Mesa 25.0's`src/c11/threads.h`, because glibc 2.43 and later declare C11's `once_flag` and `call_once` in