I work for AMD on our GPU math libraries. If you have specific complaints, I will gladly listen. I can't promise more than that, but I will read and consider whatever you write.
I'm a true believer in ROCm HIP. It has tremendous potential. I joined the company specifically because I wanted to help ensure its success... mostly by writing fast, reliable code, but also by listening to our users.
If you (or anyone else) would prefer to respond privately, my email is my HN username at gmail.
This isn't too hard to explain, I think - until recently AMD has been very strapped for cash and all their software engineering bandwidth had gone towards optimizing games (where they're still keeping up with NVIDIA where progress doesn't seem to have slowed).
This applies both to CPUs (where they have to rely on Intel for this) and toe GPUs (where their offerings are under resourced, buggy and lagging what NVidia offer).