HIP/ROCm: Macro __AMDGCN_WAVEFRONT_SIZE has been marked as deprecated

[sozforex]

Macros __AMDGCN_WAVEFRONT_SIZE and __AMDGCN_WAVEFRONT_SIZE__ has been marked as deprecated in llvm/llvm-project#112849.

Deprecation warning in ROCm 6.3.2 release notes:

The AMDGCN_WAVEFRONT_SIZE macro will be deprecated in an upcoming release. It is recommended to remove any use of this macro. For more information, see AMDGPU support.

This comment suggests an alternative ROCm/ROCm#4121 (comment)

Possibly of use - changes in OpenMP as the result of this deprecation llvm/llvm-project#113156

The relevant part of llama.cpp:
https://github.com/ggerganov/llama.cpp/blob/d80be897acdca45f8f7ea2e52c930b1d0e738a14/ggml/src/ggml-cuda/common.cuh#L179-L185

[IMbackK]

I am aware, unfortunately no replacement exits, as this value will then no longer be available as a compile time constant, which is btw very silly as the compiler has to know the wave size while compiling in the device pass.

Eventually when it is removed we will have to guess the wave size based on __gfxXXXX__ defines.

Even more eventually once amdgcn-spriv lands we will have to cease considering wave size a compile time constant at all, this however requires significant refactoring to avoid loosing performance.

For now there is nothing we can do other than do nothing.