[MKL-DNN]Bump up mkl-dnn to 0.20

[jczaja]

Those changes are updating MKL-DNN from 0.19 to 0.20.

Justification:

• Some new post operations (needed for some int8 workloads)
• better performance of sofmax & transpose ops (AVX512 platforms)
• inner product does work on 3D,4D,5D.. dimensional data so we can enable some workloads like BERT to use mkl-dnn FC more (enabling to be done after this is merged)

[luotao1]

This commit will fail on one of the internal models, which runs successfully on 0.19.

[jczaja]

@luotao1 could you please share some logs/description of failure?

[jczaja]

@luotao1 In particular I'm interested in MKLDNN_VERBOSE log

[luotao1]

The model was sent to you Dec 2018. If you don't find the model, I will ask @bingyanghuang to sync with you.

[jczaja]

@luotao1 I have updated changes with needed fix to have crash in internal workload you reported, goes away. We are testing currently those changes.

[luotao1]

@jczaja Do you mean current PR is OK for review?

[jczaja]

@luotao1 I'm sorry for being not very clear. We haven't finished testing it yet (Should be done today/tomorrow) and then those changes will be ready for review. But what I would like You to do is to test your internal workloads if possible on this branch and tell me if you see any problems . If you also let us know if there is some performance change that would be good as well.

[jczaja]

@baojun-nervana Just to let You know that we are bumping up mkl-dnn to 0.20 in this PR.

[jczaja]

@luotao1 We finished internal testing. From highlights, there is slight performance improvement on BERT inference on AVX512 platforms eg. Skylake and Cascadelake. So PR is now ready for review, if you have any data on how this PR working with your internal workloads then please share

[baojun-nervana]

@baojun-nervana Just to let You know that we are bumping up mkl-dnn to 0.20 in this PR.

Thanks, will update ngraph.

[Sand3r-]

LGTM

[jczaja]

@luotao1 Have you had a chance to test this PR on your internal workloads? If yes and positive then please consider merging

[luotao1]

It works OK on internal workloads, thanks very much!