Questions about using XNNPACK Execution Provider

[postech-sms]

Hi, all

I have questions about using XNNPACK Execution Provider. (linux-arm64 ubuntu:20.04)

I built onnxruntime including xnnpack from the v1.16.0 tag in the onnxruntime repository. (use --use_xnnpack)

Using my built files,
I checked that XNNExecutionProvider works well For the Full Precision Model.
(I mean that convolution layers are mapped into XNNExecutionProvider)

{"cat" : "Node","pid" :1047768,"tid" :1047768,"dur" :2521,"ts" :184336,"ph" : "X","name" :"XnnpackExecutionProvider_Conv_3_kernel_time","args" : {"thread_scheduling_stats" : {"main_thread": {"thread_pool_name": "session-11-intra-op", "thread_id": "281473511240736", "block_size": [], "core": 3, "Distribution": 0, "DistributionEnqueue": 0, "Run": 0, "Wait": 0, "WaitRevoke": 0}, "sub_threads": {"281472132768032": {"num_run": 0, "core": -1},"281472141222176": {"num_run": 0, "core": -1}}},"output_type_shape" : [{"float":[1,28,28,96]}],"output_size" : "301056","parameter_size" : "384","activation_size" : "602112","node_index" : "138","input_type_shape" : [{"float":[1,56,56,48]},{"float":[96]}],"provider" : "XnnpackExecutionProvider","op_name" : "Conv"}}

However, when I converted Full Precision Model to INT8 Quantization Model,
Convolution layers weren't mapped into XNNExecutionProvider
(Just mapped into CPUExecutionProvider)

{"cat" : "Node","pid" :1048801,"tid" :1048801,"dur" :1041,"ts" :276498,"ph" : "X","name" :"QLinearConv_token_1228_kernel_time","args" : {"thread_scheduling_stats" : {"main_thread": {"thread_pool_name": "session-5-intra-op", "thread_id": "281473290830880", "block_size": [1], "core": 0, "Distribution": 1, "DistributionEnqueue": 0, "Run": 964, "Wait": 14, "WaitRevoke": 0}, "sub_threads": {"281472570421536": {"num_run": 4, "core": 3},"281472410317088": {"num_run": 4, "core": 1}}},"output_type_shape" : [{"uint8":[1,16,16,96]}],"output_size" : "24576","parameter_size" : "399","activation_size" : "24576","node_index" : "298","input_type_shape" : [{"uint8":[1,16,16,96]},{"float":[]},{"uint8":[]},{"float":[1]},{"int8":[1]},{"float":[]},{"uint8":[]},{"int32":[96]}],"provider" : "CPUExecutionProvider","op_name" : "QLinearConv"}}

I confirmed in the official documentation that XNNExecutionProvider also supports QLinearConv. (ai.onnx:QLinearConv)
(https://github.com/microsoft/onnxruntime/blob/gh-pages/docs/execution-providers/Xnnpack-ExecutionProvider.md#supported-ops)
But I don't understand why XNNPackExecutionProvider doesn't work properly in my INT8 model.

This may be an issue that occurred because I am a newbie, so please understand.
Thanks.