Lunderberg commented on PR #16183: URL: https://github.com/apache/tvm/pull/16183#issuecomment-2276526876
FOUND IT! The root cause was in the type validation, but ended up being pretty easily solveable. When converting from the untyped `ArrayNode*` to the typed `Array<T>`, we validate that each `ObjectRef` in the array has the correct type `T`. However, because each element in the untyped `ArrayNode*` must still be some type of `ObjectRef`, this check is unnecessary. In the test case, this performance overhead largely occurred during the calls to `"vm.builtin.tuple_getitem"`. For each call, the validation function iterated over the entire `Array<ObjectRef>` argument. This overhead can be removed by adding a compile-time check on the type parameter's type. I've updated the conversions for both `Array<T>` and `Map<T,U>` to skip type checks that will always succeed. With this addition, the FFI update went from being 5x slower than baseline (increase from 1 ms to 5 ms) to being 30% faster than baseline (decrease from 1ms to 700 us). I'll make a PR shortly that re-applied the FFI changes, with the additional fix. -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
