masahi opened a new pull request #6455: URL: https://github.com/apache/incubator-tvm/pull/6455
I found an interesting multithreading logic bug in x86 injective schedule, for 1D loop workload. Even though this line https://github.com/apache/incubator-tvm/blob/master/python/tvm/topi/x86/injective.py#L44 would make the entire 1D loop parallel, this `split` https://github.com/apache/incubator-tvm/blob/master/python/tvm/topi/x86/injective.py#L49 would break the parallel loop and leave the outer loop after split single threaded. I found this while investigating a CPU performance issue with models coming from [hummingbird](https://github.com/microsoft/hummingbird) project. For example, [this loop](https://gist.github.com/masahi/21dc3dd9c37cb129aa3be6018c22ffa5#file-lower_20000-txt-L68) is properly parallelized, while other loops like [this](https://gist.github.com/masahi/21dc3dd9c37cb129aa3be6018c22ffa5#file-lower_20000-txt-L95) is single threaded. please review @anijain2305 @yzhliu @kevinthesun ---------------------------------------------------------------- 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: [email protected]
