hipcc does not hipify __nvvm_get_smem_pointer function about hcc HOT 3 OPEN

• @b-sumner , @yxsamliu , @emankov

from hcc.

This looks like an internal function to me, well outside the scope of the cuda programming language. Why are those "many Cuda related project" choosing such a risky direction? What is wrong with just using ordinary pointers?

from hcc.

This looks like an internal function to me, well outside the scope of the cuda programming language. Why are those "many Cuda related project" choosing such a risky direction? What is wrong with just using ordinary pointers?

@b-sumner "many Cuda related project", Yes, there a few of them, such as cutlass, apex, etc. A lot of applications using shared memory mechanism of CUDA may use it. So by your answer, I guess you mean, in AMD gpu platform, just replace this __nvvm_get_smem_pointer function with an simple type conversion from void* to uint32_t is enough? Or there are more things needed to be done other than just a simple type conversion?

from hcc.