Xeon optimizations

[sanchit-misra]

Description

This PR contains the single socket optimizations of SpMM for Xeon as mentioned in our DistGNN paper: https://arxiv.org/abs/2104.06700

Checklist

Please feel free to remove inapplicable items for your PR.

• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage
• To the my best knowledge, examples get faster or equal performance and accuracy is not affected.

Changes

Provided Xeon optimized implementations of SpMMSumCsr() and SpMMCmpCsr(). We have observed up to 4.4x speedup on the SpMM kernel without change in accuracy.

Suggestion for best performance

The optimizations can achieve better performance for dense full graphs (like Reddit) if the neighbors for each node in CSR matrix (indices) are sorted.

[dgl-bot]

To trigger regression tests:

• @dgl-bot run [instance-type] [which tests] [compare-with-branch];
  For example: @dgl-bot run g4dn.4xlarge all dmlc/master or @dgl-bot run c5.9xlarge kernel,api dmlc/master

[jermainewang]

Look like the PR has conflicts with the previous efforts of adding libxsmm into DGL. I will close this one. Please rebase and create a new PR. THanks.