FrozenGene commented on pull request #6711: URL: https://github.com/apache/incubator-tvm/pull/6711#issuecomment-716609552
> Hi @FrozenGene > Thanks for your reply! > > 1. Yes we can, but there is a issue. It's very hard to have the two quantized strategies (int8 + smull/sadalp vs int16 + smlal/smlal2) available to the auto-tuner at the same time (you already saw [this post](https://discuss.tvm.apache.org/t/quantized-models-and-legalization-pass/8253/3), I will reply there to your comment). In the case of depthwise, we only have the int16 strategy, so the issue does not arise. > 2. Nice you asked, because I am trying to prototype the idea of a "tir pass" where this tensorization (and possibly every arm tensorization in `tensor_intrin.py`) happens. My hope is to make it work more or less like `vectorize`. One of the main reasons to have this is to enable Ansor support For ansor, we could make ansor support tensorize (like TensorCore we need tensorize too). However, if we could done it in the llvm / tir, we will make ansor support it easily. If we could lift it as generic pass, I think it will bring benifit to other hardware platform too. ---------------------------------------------------------------- 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]
