Artem-B wrote: AFAICT, all small FP types in CUDA (as implemented by NVCC) had been historically handled via opaque integer types with header-only implementation of operations on them. Clang does have native fp16 (and bf16 these days?) support enabled, but it's rarely used, because it's not portable.
That's most likely the reason that libdevice has to resort to integer types to remain compatible with the existing implementation of those types. https://github.com/llvm/llvm-project/pull/174005 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
