frasercrmck wrote: > Also, just leave that thought here: A next step might be to use `native` > functions instead (when available). They should provide enough precision for > half, but be more performance. Also as the accuracy is undefined for those > functions, maybe it should be an opt-in option in the libclc (clspv would > enable it for sure).
Yeah, it would definitely need to be opt-in. Ideally we'd also implement the `half` builtins properly so they're not going through `float` and are as fast as they can be, but then _also_ allow native functions if users so wish. But that involves maths, haha. https://github.com/llvm/llvm-project/pull/98149 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
