CSCI 699 === Test approaches --- 1. C++ infunction timing. Calculate the actual time used when performing the calculation. - might be ignoring other factors 2. Js wrap timer,(or by using CL tools), calculate the time used for the whole program. - might be affected by memory allocation and releasement time. 3. --- Tasks tobe tested --- 1. AI/ML related tasks - Tensorflow lite compilation to wasm - TFlite is compile in bazel, which is super complex and is not modifiable. - It is the performance testing of the whole library. - Pytorch model -> torch script -> ONNX -> C++ runtime - Uses libtorch, and torch script to compile translate the model. - Hard to explain the actual implementation optimization or reasoning of the performance. - Same issue, performance testing of the whole pytorch library. (There are some prebuilt version) - ONNX model -> ONNC backend -> C++ runtime - Much lower level, takes time to look into the implementation of the source code. - Relatively small project, prob. is not super optimized, but not really effecting the mass usage.