Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[TIR][REFACTOR][API-CHANGE] Change Call.name to Call.op(RelayExpr) (#…
…5863) * [TIR][REFACTOR][API-CHANGE] Change Call.name(string) to Call.op(tvm::Op/RelayExpr) This PR brings a major refactor to the tir::Call structure. The current Call structure uses a string field(name) to identify the function/intrinsic being called. This approach is limited as we start to expand TIR to be more structured. In particular, we are interested in the following aspects: - Type a function and perform better compile time type checking so that we can find errors early. - Register additional properties about an operator, such as: - Whether an intrinsic can be vectorized - What is the adjoint function of the intrinsic(for tensor expression AD) - Whether the operator has side effect. - Perform specific codegen about an intrinsic if necessary. - Call into another function in the same module. The refactor changes the Call.name field to Call.op. The Call.op field has a RelayExpr type, and we can pass: - A tvm::Op which represents the corresponding intrinsic. - A tvm::GlobalVar for calling into another function in the IRModule. All the current intrinsics are migrated by registering an tvm::Op. Because the unified IR shares a single Op registry. We use the "tir" namespace for tir related intrinsics, for example bitwise and is now registered under `tir.bitwise_and`. To simplify upgrade, we introduce a `tir.call_extern` intrinsic that allows us to call into arbitary external function without type checking. However, we should move towards more type checked variants in the system. Under the new op design. We should no longer try to pattern match all the specific intrincis. Instead, we should rely on attr of each Op to do transformation. For example, the vectorization pass depends on the TVectorizable property of the op, which can be registered independently. In this way, we can still grow the number of intrinsics when necessary without having to change all the passes. The same rule applies for tensor expression AD. Currently we are performing AD by pattern match on operators like exp, sin, cos. We should instead change to the ajoint registeration mechanism like those in relay. Followup refactors need to be performed, including: - Fold the Call.call_type into operator's attribute. - Enrich the operator registry information - Refactor passes(e.g. AD, intrin lowering) to use the attribute based transformation * Fix nms * Fix remaining testcase * Address review comment
- Loading branch information