This is a great disscussion here. Actually, we are supporting a DSA with TVM, let me share my practice.
1, We only re-use some of tvm relay or tir passes, less than 10 passes, such storage flatten, we don't need most of tvm passes, keep them in our flow means wasting compilation time. 2, We develop passes, and enhance tvm passes for our target, such stroage rewrite. 3, We develop a hybrid ir on tir, in which we can do unified memory allocation and schedule insnstruction across operators. So we are excited to see tvm can do these in main line. What we want to have: 1, customzation compilation flow for both relay and tir flow. 2, unified ir to view both graph and tir, supporting inter/intra pass developing. -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/apache/tvm-rfcs/pull/10#issuecomment-904635448
