[pjmlp]

I think the main issue is that most research is done in C++ and Fortran.

Both are not supported by OpenCL, which is one of the reasons many researchers go CUDA instead.

Khronos finally got out of their C only APIs mentality with the upcoming revision of OpenCL, but apparently AMD might have felt it is already too late to make developers use something SPIR based, with most vendors already targeting PTX.

[slavik81]

Why would it be too late for something SPIR-based? Vulkan and OpenCL 2.1 will be compiling to SPIR-V.

[pjmlp]

Existing code bases and knowledge with the current tools.

You will need C++ and Fortran compilers able to target SPIR with the same set of language features, debugging tools and libraries that CUDA has available today.