Every GPU vendor knows that actually supporting a reimplementation of CUDA for their hardware immediately signs them up for two problems:
1. Protracted lawsuits with the (now) second-most-valuable company in the world.
2. Any significant customers being loudly reminded by Nvidia that the EULA for CUDA tools prohibits their use on non-Nvidia hardware.
The problem is not the technical challenge of reimplementing the CUDA API or tools. The problem is the users want CUDA, not something that looks and behaves like CUDA but requires them to load a different set of libraries.
1. Protracted lawsuits with the (now) second-most-valuable company in the world. 2. Any significant customers being loudly reminded by Nvidia that the EULA for CUDA tools prohibits their use on non-Nvidia hardware.
The problem is not the technical challenge of reimplementing the CUDA API or tools. The problem is the users want CUDA, not something that looks and behaves like CUDA but requires them to load a different set of libraries.