merrymercy edited a comment on pull request #6040: URL: https://github.com/apache/incubator-tvm/pull/6040#issuecomment-659141089
I added a python version. It handles parallel and vectorization generally for most input shapes. (However, in my test, the vectorization part is not useful. I guess llvm can handle it) I tested in with BERT (seq_len=128, batch=1) on c5.9xlarge. It reduces the latency from 19.0 ms to 18.7 ms. My questions: 1. Why do we have both c++ version and python version? Should we just delete the python version? 2. I see the same name (topi.x86.schedule_injective) appears in both python and c++. Which one overwrite the other one? It is not clear in the code. - python https://github.com/apache/incubator-tvm/blob/71b48adf64ffe977c0c89a5fba7b6051eab891d7/topi/python/topi/x86/injective.py#L94 - c++ https://github.com/apache/incubator-tvm/blob/dff715a54ee0f02f27b4f4efde08c77e86eff2d2/topi/src/schedule.cc#L99 ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: us...@infra.apache.org