The RDNA docs are just the tip of the iceberg.. in order to get to the compute engines (what you actually care about) you have to go through 3 layers of user-space cruft, a kernel driver, and then a firmware layer running half a dozen separate components of the GPU that "manage" the RDNA compute engines. Apparently, these components run on at least 3 different ISAs (some ARM, some F32, and some on RS64) where 2 of them aren't really documented at all. All of this is what he's trying to bypass and talk to the RDNA cores as directly as possible. Modern consumer GPUs are complex beasts.