resenet50 ONNX int8 model inference run in ONNXRUNTIME vs IREE-COMPILED CPU run differ

[kumardeepakamd]

There is 11% match only between ONNXRUNTIME run and IREE compiled inference for VAI quantized int8 model of resenet50.

To reproduce the issue:

git clone https://github.com/nod-ai/SHARK-TestSuite.git
cd e2eshark
(You may need to read the https://github.com/nod-ai/SHARK-TestSuite/blob/main/e2eshark/README.md and make sure you have the right pip install etc.)

Replace the path names for hfhome, -c and -i to point to your Hugging Face Home, Torch MLIR build and IREE builds respectively and run as:

python ./run.py --hfhome /proj/gdba/kumar/HF_HOME -c ../../torch-mlir/build -i ../../mainireee/iree-build --torchtolinalg --tests onnx/models/resnet50_vaiq_int8

cd test-run/onnx/models/resnet50_vaiq_int8 and examine the logs, you can 'cat commands.log' and rerun any step you like. You will see that failedinference.log shows you difference in results