Hi @akote123 , could you share the original verbose log, thanks!

Hi @huiyan2021, I have uploaded here vit_intel_log.zip

I guess the reason is that _mklops are logged by <<, where the outputs are fully buffered if they are redirected to a file, while oneDNN verbose always flush stdout immediately: https://github.com/search?q=repo%3Aoneapi-src%2FoneDNN+fflush&type=code

@huiyan2021 , one more observation what I found is some ops are common_runtime/eager/execute.cc and some are in common_runtime/executor.cc , here I am not able to understand why there is two path for execution

Hi @akote123,

from the log I can see you are using xla, so for ops can not be jitted, they come to common_runtime/eager/execute.cc, otherwise they come to common_runtime/executor.cc

also, suggest that you use trace viewer to trace executions.

@huiyan2021 , do we have file location tensorflow source code where checking happen whether to go common_runtime/eager/execute.cc or the other one.

@akote123 , you can refer to this article: https://whatdhack.medium.com/tensorflow-graph-graphdef-grappler-xla-mlir-llvm-etc-615191e96ebc, see XLA Flow part and call stack

@huiyan2021 , In tensorflow the single model can go in both XLA and oneDNN or is it like either it will use XLA or oneDNN only

Both. There may be different scenarios:

1. Some parts of the model go in XLA path, some parts go in oneDNN path.
2. Intel recently submitted a pilot PR to accelerate XLA’s Dot op with oneDNN.

You can refer to this RFC: https://docs.google.com/document/d/1ZzMcrjxITJeN2IjjgbzUjHh-4W1YgDUus3j25Dvn9ng/edit

@huiyan2021 ,Thank you for the pointers I will got through it .
Actually for pretrained models how we can enable XLA both for inference and transfer learning

same as training, you can refer to https://www.tensorflow.org/xla#enable_xla_for_tensorflow_models