Much longer calculation time using dnnl

[niclas-b]

Hi,

I tested my own model using the onnxruntime_perf_test.exe once with dnnl and once without. Although the results show, that the dnnl runs three times faster during inference while set up time is more or less the same, the tool needs very long to finish the experiment, while in the case of not using dnnl I get the result immediately. What is the tool doing in the meantime when this overhead is not shown in the results?
My idea was that it needs to compile the model first, but then if I run a dnnl session by script, this overhead does not exist. In this case it's as fast as running the model without dnnl.

Best Regards
Niclas