search

apache / tvm

Open deep learning compiler stack for cpu, gpu and specialized accelerators
https://tvm.apache.org/
Apache License 2.0
11.58k stars 3.43k forks source link

what about the support for third-part accelerators? #4110

Closed meiniangpp closed 4 years ago

meiniangpp commented 4 years ago

Hi, it's a very impresive tool. And I have 2 questions.

  1. As for the auto-tuning, is it tunning the structure of the dnn, or tunning the scheduling and/or compiling on the target platform?
  2. Does it support for third-part accelerators? Such as some TPU like accelerators of our own?

Thank you a lot for your time!

comaniac commented 4 years ago

Hi, for your questions:

  1. The AutoTVM is for performance tuning. It tunes scheduling for each operators in the network as well as the graph level data layout for the given platform. The architecture tuning, which refers to neural architecture search, is not supported yet in the current TVM.

  2. TVM currently does not really support third party accelerators but we do have an ongoing plan for it.

p.s. it’s suggested to post general questions like these to the discussion forum (discuss.tvm.ai) instead of GitHub issue.

meiniangpp commented 4 years ago

Thank you very much for your time. And next time I would post questions like these to the discussion forum. Best wishes!