AM68A: AM68A

Hi,

I have some questions about TI PSDK9.0 Quantization.

I am using edgeai-tidl-tools to create TIDL-RT.

• edgeai-tidl-tools : https://github.com/TexasInstruments/edgeai-tidl-tools
• tvm-compilation-onnx : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/examples/osrt_python/tvm_dlr/tvm_compilation_onnx_example.py

TIDL-RT binary files are created by using tvm-compilation-onnx.py. I can find some parameters that looks like for quantization such as dataQ, roundBits, min/maxValue, tensorScale in bin.svg file.

Q1.

Can I manually change the quantization parameters?

If there is a documentation for parameters, Can I receive it?

Q2.

Are asymmetric, non-power of 2 scale, per-channel quantization supported in psdk9.0?

These contents have not been updated in a edgeai-tidl-tools document.

• https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_quantization.md#a4-futureplanned-improvements:~:text=A.4-,Future/Planned%20Improvements,-The%20following%20options

 
I found parameters [quantization_scale_type, output_feature_16bit_names_list, params_16bit_names_list, quantization_scale_type, channel_wise_quantization] in delegate_options in onnxrt_ep.py
Can I have documents for those parameters?
-  github.com/.../onnxrt_ep.py

Q3.

Is there a tool available for calculating the similarity between fp32 model and int8 model created by tidl quantization?

Thanks,
YoungHeon