NNAPI acceleration of random forest

[OrdevAg]

I created a model in scikit-learn using Random Forest Classifier and converted it to ONNX using sk2onnx. I want to run this model for a large set of samples on Android smartphone using C++ onnx runtime. In my C++ code I enable Nnapi provider:

uint32_t nnapi_flags = NNAPI_FLAG_USE_FP16 | NNAPI_FLAG_CPU_DISABLED;
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_Nnapi(sessionOptions, nnapi_flags));
Ort::Session session(env, model_name, sessionOptions);

I checked that onnx runtime detects Nnapi provider (using Ort::GetAvailableProviders()), it finds NNAPIExecutionProvider. Unfortunately there is no any acceleration comparing to CPU-only runs. Is there a way to accelerate the inference of Random Forest model?