xenos-linux-gpu

Hardware-accelerated OpenGL ES 2.0 and an accelerated X server on the Xbox 360 GPU (ATI Xenos), under Linux.

This repo holds everything from a two-day bring-up on a softmodded Xbox 360 (Trinity slim, Peer Pressure softmod, Xell): a kernel module that drives the Xenos command processor and implements enough of the msm DRM interface for Mesa, a Xenos variant of Mesa's freedreno a2xx driver, a big-endian patch for Xorg's glamor, the Khronos conformance harness we used to find bugs, and the tools that let a machine with no GPU reset debug its own hangs without a human pressing buttons.

The live scanout buffer, read from memory and detiled on the host: Xorg with glamor, all rendering on the Xenos GPU, scanned out directly from the GPU's tiled render target. xclock, xeyes, xlogo and xterm, stipple root window, cursor.

Status (29 October 2026)

Kernel	Linux 6.18.58 + Free60 xenon patch (plus sis190 TX fix, VMX enabled but not advertised), netbooted by Xell, Debian unstable ppc64 on NFS
GPU driver	xegpu.ko: CP microcode and ring, fences on interrupts, boot-time contiguous pool (80 MB, one range), KMS display, msm UAPI subset with a command checker
GL	Mesa freedreno a2xx with a Xenos variant (gpu_id 290), GLES 2.0 and desktop OpenGL 2.0 (EGL, GLX), in X windows through DRI3
Conformance	dEQP-GLES2 must-pass list, all 17,165 cases: 99.9% of executed cases pass (17,063 of 17,075), no GPU hangs, no crashes (conformance/results/full41-report.txt). 12 cases still fail, listed in docs/conformance.md
X	Xorg 21.1 modesetting + glamor (GLES2), depth 24/32 accelerated, no hangs with xclock/xeyes/xlogo/xterm
Performance	Four rounds (docs/performance.md): fills, text, composites, large copies, scrolls and uploads beat CPU rendering (rect10 2.9M/s vs 1.7M, aa10text 839k vs 290k, scroll500 5,480 vs 904); tiny copies, tiny uploads and readback still lose. glmark2-es2: all 32 scenes run, 34-898 FPS at 800x600 (conformance/results/glmark2-2026-10-28.txt)
Desktop	IceWM on glamor by default (Xfce works but 458 MB is tight), GLES and GLX apps on the GPU in X windows. Tried: NetSurf (Wikipedia with images), neverball and chromium-bsu (GL 1.x/2.0 games, same output as Mesa's software renderer)
Desktop GL	OpenGL 2.0 / GLSL 1.20; piglit GL 2.0 subset: 90.4% of executed tests pass (1,031 of 1,140), no GPU hangs (docs/conformance.md)
Not done	GPU reset for real wedges (a CP-only reset works, a runaway shader still needs a power cycle), display modes other than 720p over HDMI, ETC1, 4 KB-page kernel (does not boot yet), GL 2.1+, Vulkan (never)

Nothing here writes to the console's NAND. The only change to the softmod is one byte in a file on the hard disk (see docs/setup.md).

Caveats

Read these before reusing anything here.

• This is an AI-driven experiment. Claude (Opus) did the work in a loop over about three days, with a human setting direction. Treat it as a proof that the hardware can do this, not as reviewed code.
• The kernel side is a standalone shim, not the upstream msm driver. xegpu.ko is libxenon's GPU code moved into a kernel module, plus a fake msm UAPI so Mesa's freedreno accepts it. People who know this hardware think upstream drivers/gpu/drm/msm (which already handles Adreno 2xx) could drive Xenos with a few patches. That is very likely the better design; we never evaluated it.
• The first AltiVec kernel patch was wrong, and has been replaced. It advertised AltiVec to user space, but the Xenon's VMX128 lacks parts of AltiVec (the integer multiply and sum family), and Free60 leaves it off for exactly that reason: software that sees it advertised picks vector code that dies with SIGILL (pixman, libjpeg-turbo; NetSurf died on the first JPEG of a web page). kernel/patches/0002-xenon-vmx-not-advertised.patch keeps VMX enabled for code built with AltiVec (Debian's libLLVM) but hides it from AT_HWCAP and /proc/cpuinfo, so libraries choosing at run time stay on their scalar paths.
• Several "hardware rules" are measurements on one console (a Trinity slim), not documentation. Where a rule says something is impossible, read it as "we didn't find a way".

What was new

Most of the work was finding out where Xenos differs from the Adreno 200 that freedreno targets. They share an ISA and most of a register map, which made freedreno the right starting point and also made every difference silent: the GPU just stops. docs/hardware-rules.md lists each rule with the evidence. The ones that cost the most time:

• Pixel shaders get 16 GPRs after libxenon's Xe_Init. Any shader using a 17th register wedges the GPU. The kernel now splits the pool 64/64 through the ring the way Xe_Init does; a plain MMIO write to SQ_GPR_MANAGEMENT does nothing.
• GL's first submit wedged the GPU unless libxenon's init stream had run since the last CP init. The kernel now replays it before the first msm submit.
• ALU constants must be written as whole 16-constant blocks, fetch constants as whole 6-dword groups.
• The vertex shader must export every parameter slot it allocates, position first.
• Render targets smaller than 8x8 pixels wedge the GPU.
• COHER_STATUS_HOST bit 25 invalidates the vertex cache; bit 24 (what libxenon uses) only covers textures.
• Vertex fetch stride and offset are in dwords; the index DMA size is in 16-bit words; there are no 8-bit indices; primitive type 7 is not a line loop.
• EVENT_WRITE_SHD with CACHE_FLUSH_TS wedges the memory hub. Fences are WAIT_UNTIL idle + MEM_WRITE.
• Indirect buffers nest two deep at most.
• The Xenon CPU's VMX128 unit lacks the AltiVec integer multiply family, and the core lacks ldbrx/stdbrx, which -mcpu=cell emits. Build user space with -mcpu=970 -mtune=cell -mno-altivec.

How it was debugged

We have no working software reset for the Xenos: libxenon never implemented one, and the RBBM soft-reset sequence we tried hung the console (hardware rule 21). That's a gap in our knowledge, not a known hardware limit. So a wedged GPU stays wedged until the console loses power, and a bad memory write takes the whole console with it. We built an IR blaster (Arduino + IR LED) and patched one byte of the softmod so that a Windows Media Center remote's power code boots straight into Xell and Linux. A host script can now power-cycle the console in about 80 seconds, so a test harness can run thousands of GL tests, survive hangs, and attribute each hang to a test. docs/methodology.md covers that, the replay tools, the Xenia-based linters and the footguns we hit on the way (docs/footguns.md).

Layout

kernel/xegpu/        xegpu.ko: CP, ring, fences, pool, KMS display (xenos_drm.c), msm UAPI (msm_compat.c)
kernel/include/      xegpu_uapi.h
kernel/xenon_idle/   low-priority idle loop for the Xenon SMT threads (1.85x single-thread speed)
kernel/patches/      Free60 kernel fixes: sis190 TX wake race; Xenon VMX enabled but not advertised
userspace/           libxenon's xe.c/edram.c ported to Linux user space, xeinit (EDRAM training), tests, replay tool
mesa/patches/        git format-patch series on Mesa (base ab1502f), cross file
xorg/                glamor big-endian GLES patch (xorg-server 21.1.25), xcopytest (copy/PutImage checker)
gltest/              small GL probes, each aimed at one hypothesis
conformance/         dEQP-GLES2 harness with hang attribution and IR power cycling; dEQP big-endian patch
tools/               rd dump replay/bisect (rdstep), Xenia-table linters (xenoslint, xenosfields), decoders
irblaster/           Arduino sketch and host scripts for hands-off power cycling
softmod/             FATX in-place editor, SMC power-on reason reader
scripts/             bring-up scripts run on the console
docs/                setup, hardware rules, methodology, footguns, conformance, glamor, timeline

Docs

• docs/setup.md: netboot, NFS root, softmod patch, building everything
• docs/apps.md: browsing, video and audio, games: what runs, how to start it, and the memory limits
• docs/hardware-rules.md: Xenos vs Adreno 200, with evidence
• docs/methodology.md: how to debug a GPU that can't be reset
• docs/footguns.md: things that cost us hours
• docs/conformance.md: dEQP pipeline and results
• docs/glamor.md: getting X accelerated
• docs/performance.md: x11perf numbers, profiling without perf events, what changed
• docs/timeline.md: what happened when
• docs/isa-a2xx-vs-xenos.md, docs/path-c-plan.md: early design notes

Credits

Built on Free60 (kernel patch, libxenon, Xell), the Xenia emulator's register and ISA documentation (src/xenia/gpu), Mesa's freedreno a2xx driver, Khronos VK-GL-CTS, and Peer Pressure. The work was done by an AI coding agent (Claude) driving a real console, with a human doing the soldering-iron-free hardware bits and the steering.

Upstream

Nothing here has been submitted to Free60, Mesa, the X server or Khronos, and that's deliberate: the work hasn't been reviewed by a human who understands it well enough to stand behind it, and unreviewed AI-written patches are a burden on maintainers. Use it, learn from it, but please don't forward these patches upstream on our behalf.

License

See LICENSE.md. Kernel code and the libxenon port are GPL-2.0, Mesa patches are MIT like Mesa, the dEQP patch is Apache-2.0 like VK-GL-CTS, docs are CC BY 4.0.