Roadmap for supporting more operations with QS8
shanumante-sc opened this issue · comments
shanumante-sc commented
Currently XNNPACK only supports a subset of quantized operations for QS8 inference (conv, fully connected, add, global avg pool); a much larger set of operators is supported for QU8. Are there plans to implement missing QS8 operators?
For our use case specifically, it would be great to have avg/max pooling; deconvolution; sigmoid/softmax/clamp operators supported with QS8 inference.
Marat Dukhan commented
In the near term, QS8 support will focus on missing operators from MobileNet v3, i.e. mul and hard swish. If you'd like to have other operators, contributions are welcome!