SCALE is not a "translation layer", it's a full source-to-target compiler from CUDA-like C++ code to AMD GPUs.

See this part of the documentation for more details regarding warp sizes: https://docs.scale-lang.com/manual/language-extensions/#impr...