There is a new implementation of roctx in ROCprofiler-SDK: https://rocm.docs.amd.com/projects/rocprofiler-sdk/en/latest/how-to/using-rocprofv3.html#marker-trace The new header is [`rocprofiler-sdk-roctx/roctx.h`](https://github.com/ROCm/rocprofiler-sdk/blob/amd-staging/source/include/rocprofiler-sdk-roctx/roctx.h) and the new runtime library is `librocprofiler-sdk-roctx.so`. It is supposed to have less runtime overhead.