Yes but if it is patented that doesn't mean it is secret -- in fact the opposite. I can copy the code but I can't sell a product using it without getting a licence or risk being sued. Obviously I can't licence it as GPLv3. But I don't see why the driver implementation should be so closed.
Because you don't hide the driver to hide what's in your patent. You hide the driver to hide code that violates someone else's patent.
'Patent minefield' means there are so many patents for so many basic things it's impossible to do anything without violating something. You'll never know you've violated a patent until your competitor sues you. A lot of these patents are probably invalid, but the fights in court will take years, judgement will sometimes happen by someone without know how, and losing 1 fight will cause you severe damage.
The 'solution' is trying to hide all your violations by closing the driver. Reverse engineering costs a lot, so if someone is doing it to find your patent violations, you build your own patent war chest and cross license with them - they have engineers so they probably produce something that has violations for your patents.
Ah, that makes total sense. Makes you wonder how they avoided this situation with RISC-V (they say via identification of prior art), but graphics does seem to be worse patent-wise (though ARM is notoriously litigious).
Having spent a few years focused on this (and other semiconductor enablement-related problems) my guess is they have evaluated their own patent portfolios and the risk of their implementations triggering a war, and have concluded it was worth doing. Note that Intel's GPUs are relatively simple, so AMD would seem to be bolder here, but I haven't dug into the detail to see how they have limited their exposure.
I find Intel is pretty solid about open sourcing everything they provide for Linux that runs on the host — it's a pity they haven't applied the same policy to the ME.