Created by: luotao1

The test machine is: E5-2620 v2, which only have AVX Instruction set.

When I inference resnet50 model: https://github.com/PaddlePaddle/Paddle/compare/develop...luotao1:resnet50?expand=1

The commit id is: https://github.com/PaddlePaddle/Paddle/commit/643b6faa0ced3304d5531dfa6a87a7fb39a54e0f

I run ctest -R test_analyzer_resnet50 -V. following Test fails:

// Compare result of NativeConfig and AnalysisConfig
TEST(Analyzer_resnet50, compare) 

The error is that there is > 0.001 diff for all result data.

95: The difference between pdata_ref[j] and pdata[j] is 0.00164794921875, which exceeds 1e-3, where
95: pdata_ref[j] evaluates to 451.11993408203125,
95: pdata[j] evaluates to 451.1182861328125, and
95: 1e-3 evaluates to 0.001.

When FLAGS_with_mkldnn=true:

• NativeConfig: mkl
• AnalysisConfig: mkldnn. And to avoid fc_fuse_pass, I exclude this fc_fuse_pass as well. Thus, the only diff is between mkl op and mkldnn op.

However, as test_analyzer_ocr runs sucessfully. We compare the ops in ocr model and resnet50 model. The difference is batch_norm op, which doesn't exist in ocr model.

Besides I also merge #13486 to see whether conv bn fuse pass runs sucessfully. But both I set cfg->_use_mkldnn = true and cfg->_use_mkldnn = false, there is still diff.

Could you help see it?