Open jhuber6 opened 1 month ago
@AnastasiaStulova I would greatly appreciate some pointers on where to start looking to enable this for the C/C++ languages. We want this behavior when using the attributes directly in the C++ based languages. I tried hard-coding the OpenCL version to 2.0 but couldn't find exactly where this check lives.
The OpenCL 2.0 rules allow us to implicitly convert specific address spaces into generic address spaces. This is also how languages like CUDA or OpenMP work. Using C/C++ for targeting the GPU should also allow this, but it is currently unnecessarily restrictive. See the example in https://godbolt.org/z/oa5Tob8v6.