@junrushao1994 so the concern is embedding e.g. the output of `NDArray::repr` into TIR text format, correct? That makes sense to me. It seems like small tir.constant (e.g. perhaps we could constraint this as: tir.constant with shape that can be squeezed to rank 2 and, when doing so, can reasonably be print each row on a single ilne) could be inlined. this would make it easier to analyze optimized kernels.
other things it seems like we need to place in an external `.params` somehow and refer to it from text format. That is more problematic as it suggests API changes and semantic changes in TVMScript: - how does one supply the data file containing said params? - how are they named? shouldn't those names be unique across the IRModule? - what if someone prints a TIR PrimFunc--currently we allow anonymous Vars so I presume we do not allow the same with such tir.constant? - how does this impact the signature of `tvm::runtime::IRModule::Add()`? None of these questions are addressed here and I think we need to if we are adopting the IRModule metadata approach. -- 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/22#issuecomment-925226956
