Created by: wojtuss

When applying the Qat2Int8MkldnnPass to the Qat1 ResNet50/101 model, INT8 accuracy of the first iteration is correct, but iterations > 1 give broken (usually 0.0) accuracy. FP32 accuracy of all the iterations on the same data is correct. INT8 accuracy for MobileNetV1/2 is fine. This suggest some oneDNN (MKL-DNN, DNNL) caching problem, possibly with conv2d's residual data.

System information -PaddlePaddle version: feba1318 -CPU -oneDNN: from Paddle -OS Platform: Ubuntu 16.04.5 LTS -Python version: from Paddle

To Reproduce

• build Paddle with tests,
• run the command in the build directory: ctest -R test_qat2_int8_resnet50_mkldnn -V
• get the test command from the beginning of the output of the above command and change the --qat_model option value to a path to the ResNet50 QAT1 model (build/third_party/inference_demo/qat/ResNet50_QAT/model/)
• run the modified command.

Describe your current behavior INT8 accuracy of the first iteration is correct, the second iteration gives wrong accuracy (close to 0.0).