Fix mkldnn backend when using naive engine

[ZhennanQin]

Description

This can fix #15078 and resolve all mkldnn related failures mentioned in #15005.

@pengzhao-intel @TaoLv @eric-haibin-lin @zheng-da

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http:https://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$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]

LGTM

[zheng-da]

Could you explain why this modification can fix the problem?

[ZhennanQin]

@zheng-da Sure. The root cause is, PushAsync works differently between threaded_engine and naive_engine.
For threaded_engine, PushAsync really works asynchronously, which will push MKLDNNDataReorderAsync to the tail of engine task queue. This is the expected behavior.
But for naive_engine, PushAsync will execute weight reorder immediately, just as a synchronous operation. I'm not sure if this is expected, but this isn't as asynchronous as function name indicates.

Now let's look at conv weight case,

weight_mem = GetWeights(weight, fwd->fwd_pd.weights_primitive_desc(), param.conv_param.num_group);

At this time, weight is still in default format, so a reorder is registered inside mkldnn stream, but not submitted. And then,

weight.MKLDNNDataReorderAsync(fwd->fwd_pd.weights_primitive_desc());

When naive_engine is activated, This will inplace reorder weight into mkldnn format immediately. So weight is reordered twice before mkldnn conv kernel launch, causing incorrect result.

The fix is quite simple, change the sequence of above code, to ensure

weight.MKLDNNDataReorderAsync(fwd->fwd_pd.weights_primitive_desc());

happens prior to

weight_mem = GetWeights(weight, fwd->fwd_pd.weights_primitive_desc(), param.conv_param.num_group);

So for GetWeights, the weight is already reordered, no extra reorder will be registered in mkldnn stream.

Do we have plan to adjust naive engine PushAsync behavior to be as same as threaded_engine? I'm worried about this will cause trouble again for other developer.

[abhinavs95]

@mxnet-label-bot add [MKLDNN]

[pengzhao-intel]

@zheng-da @eric-haibin-lin please help take a review :)

[eric-haibin-lin]

LGTM. Shall we add some documentation in the code, too?

[szha]

@ZhennanQin NaiveEngine's PushAsync was intentionally synchronous. Other code should not make any assumption about execution order when using async interface of any engine.

[szha]

In addition to the docs, should we add test for it?

[eric-haibin-lin]

We need a separate test which set naive engine env var before mxnet is imported

[ZhennanQin]

@szha @eric-haibin-lin We'd better to create naive engine CI for all UT, not only for some mkldnn tests. Naive engine should have the ability to pass all UT.

[pengzhao-intel]

I am going to merge this PR to fix the problem inside MKLDNN implementation under the naive engine.
We will submit the separated PR to enable naive engine in CI @juliusshufan