That makes it worse? The Imagination crap is shipped on Android phones, too, so there are at least vendor kernel drops to work from.

The odds of Apple releasing a development ARM board is probably roughly equal to that of them releasing complete documentation and linux kernel driver support to go along with it.

The vendor kernels are worthless, tainted, piles of crap. Vendors rarely, if ever, maintain them once they are posted publicly. The GPU still requires proprietary blobs to be even remotely useful.

Of course. But they contain non-zero amounts of information on the chip so you can do a proper implementation or, much less ambitious, update it for newer kernel versions.

Not in my experience. The code in the kernel is usually just a shim for loading the blobs that do the real work.

All the reverse-engineered GPU drivers we have now (freedreno, etnaviv, nouveau, lima) started out with intercepting kernel ioctls, being able to observe the blobs at work. Many of them at one point ran with the vendor kernel interface but open-source userspace.

In many ways the kernel part of a GPU is arguably the least important part and it's the userland where the shader compiler sits and produces the magic, but if you start from nothing it's very useful to have a working kernel interface that initializes the hardware and accepts command streams to run.

This is marketing hype.

Apple’s commits to the LLVM (where the shader gets compiled into GPU executable code) states the A11 GPU target is “PowerVR Rogue” ISA. Which is identical to how it compiles for the A10.

So it’s still using foreign IP for the instruction set. Maybe Apple moved things around/optimized internals? Get better path optimization, or improve IPC?

But it’s weird they didn’t add optimizations specifically for that A11. So an A10/A11 GPU just run the same opcodes.

Citation for those commits?

Apple's GPU is designed just down the road from Imaginations HQ. Guess where the designers previously worked...