Add primitive cache for MKL-DNN sum(elemwise_add operator

[ciyongch]

Description

This PR cached the created sum primitive to reduce the creation overhead, which is helpful to improve the performance of current MKL-DNN sum (elemwise_add) operator.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• Feature1, tests, (and when applicable, API doc)
• Feature2, tests, (and when applicable, API doc)

Comments

• If this change is a backward incompatible change, why must this change be made.
• Interesting edge cases to note here

@pengzhao-intel @TaoLv

[pengzhao-intel]

Thanks for the improvements. Could you share the performance data?

[vandanavk]

@mxnet-label-bot add [pr-awaiting-review, MKLDNN]

[ciyongch]

Here's some performance data I collected based on a customer case, primitive cached brings ~1.6x speedup for elemwise_add operator, which help to reduce the overall percentage:

1) Don't cached sum primitive (current code base)

Op name	total time(ms)	ave time (ms)	ncalls	percentage
elemwise_add	382.533	0.4	920	24.96%

2) cached sum primitive

Op name	total time(ms)	ave time (ms)	ncalls	percentage
elemwise_add	230.243	0.25	920	18.29%

[pengzhao-intel]

Why don't put this function into mkldnn_base-inl.h?

[ciyongch]

Currently, some MKLDNN ops supports ndim [1, 4], while some ops still doesn't support ndim=3.
For now, there would be several such support check functions for different ops.
But indeed, we can combine all the similar function when all the ops are finalized.

[pengzhao-intel]

This is the common approach for MKL-DNN integration so it looks good.

[pengzhao-intel]

Merging now. Thanks for your contribution.