leeexyz commented on pull request #7879: URL: https://github.com/apache/tvm/pull/7879#issuecomment-823822447
What I need to explain more is in our case, we use compute to define the HW intrinsic, it is may as the same as Operators, such as element-wise. But it is not the same compute (I mean the same addresses of all nodes), instead of the concept. :) For example, an Add op has 4 dims (i0, i1, i2, i3) and the three innermost (i1, i2, i3) axes will be split, then after reorder, the region of the inner axes (i1.i, i2.i, i3.i) will be tensorized to an Intrin_Add. In this situation, Add op and the Intrin_Add share the compute. So, in the reusing case, it is more natural to `call_extern`, I think. Anyway, I think it is better to fix this issue by updating LHS also. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org
