jwfromm opened a new pull request #8620: URL: https://github.com/apache/tvm/pull/8620
This PR adds support for non-scalar zero point values in qnn conv2d operators and also allows the kernel zero points to be channel-wise. This is a needed change to support ONNX's `ConvInteger` nodes which typically treat zero points as expressions and are often generated using OnnxRuntimes quantization feature, that produces channel wise zero points. Although the rest of the qnn framework doesn't yet support non constant zero points, this is a good start that improves our onnx coverage considerably. I also found that although qnn supported lowering uint8 convolution and dense to cuda, the dp4a instruction actually only supports int8 datatypes, an error exposed by the onnx frontend tests. I added some legalization logic to convert uint8 to int8 when the target is cuda. -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
