A GPU thread binding question

Just found some old discussion on the same topic: https://github.com/apache/incubator-tvm/issues/756

At the end of the thread, Tianqi briefly explained why normalization don’t rebase bound paths. My understanding is that the workaround at that time was to ask user to do an explicit rebase. However, I don’t see rebase is exposed as a schedule primitive today. How to rebase explicitly today, or solve this issue w/o rebase?

After almost two years, has the situation of thread binding semantics definition changed, so that my attempt shown in the previous comment may be applicable?

@tqchen, @derisavi-huawei, can you shed some light upon this?

2 Likes