We had a meeting with @Hzfengsy today. We discussed the difference and similarity of our solutions. They are different in the front-end: our solution tries to make it as transparent as possible to make it easy-using while #4095 provides more controllability to the user (schedule developer). They are actually targeting different users, so we think both solutions can co-exist. But we both agreed that the intrinsics in the back-end should combine. As to the fragment allocation, we are OK to change from new_expr to the way of introducing new scopes, but currently the new scope introduced in #4052 is not enough for the codegen of fragment allocation if it's extended to support different warp tile sizes and data layouts (col_major/row_major). One possible but not so elegant solution we proposed is to extend the scopes to also include tile size and data layout. @Hzfengsy is also trying to figure out a solution here. We will have more discussions on this. cc @tqchen
-- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/4105#issuecomment-542032766
