[NNAPI EP] add uint8 support for Transpose/Concat/Maxpool, add support of QLinearSigmoid #6534
Conversation
| // Not running using quantized input | ||
| if (input_type == ONNX_NAMESPACE::TensorProto_DataType_FLOAT) | ||
| return true; | ||
|
|
There was a problem hiding this comment.
Does this mean quantization isn't relevent here, and maybe this check be outside of IsInternalQuantizationSupported?
onnxruntime/core/providers/nnapi/nnapi_builtin/builders/helper.cc
Outdated
Show resolved
Hide resolved
onnxruntime/core/providers/nnapi/nnapi_builtin/builders/op_support_checker.cc
Show resolved
Hide resolved
onnxruntime/core/providers/nnapi/nnapi_builtin/builders/op_support_checker.cc
Show resolved
Hide resolved
onnxruntime/core/providers/nnapi/nnapi_builtin/builders/op_support_checker.cc
Show resolved
Hide resolved
| // This should not happen, but if it happens make sure this will require an impossible version | ||
| if (!GetType(*node.InputDefs()[0], input_type)) | ||
| return std::numeric_limits<int32_t>::max(); | ||
|
|
There was a problem hiding this comment.
If it should never happen (type inferencing should always have populated this value), why would we not throw?
There was a problem hiding this comment.
Want to be on the safe side, this function does not return status, so don't want this crash the runtime
|
General comment. When you have so much implicit knowledge it can be hard to recognize the lines of code that may need a quick explanation, but it would be very helpful to someone new to the whole setup if there were a few more comments explaining the 'why' behind various things. |
There will be separated work to add documents for both NNAPI and CoreML EP, |
Description: [NNAPI EP] add uint8 support for Transpose/Concat/Maxpool, add support of QLinearSigmoid
Motivation and Context