Improve binary op kernel with MKL

[eric-haibin-lin]

Binary op kernel is used for distributed kvstore at server side for gradient aggregation: https://github.com/apache/incubator-mxnet/blob/073c898e6abf415f98813acd52f11145962c96a0/src/kvstore/kvstore_dist_server.h#L732

but it is not using vectorized vsadd in MKL in its implementation:

https://github.com/apache/incubator-mxnet/blob/8df946978d68713237fd520605adc22729d241aa/src/ndarray/ndarray.cc#L890-L903

@PatricZhao @TaoLv

[mxnet-label-bot]

Hey, this is the MXNet Label Bot.
Thank you for submitting the issue! I will try and suggest some labels so that the appropriate MXNet community members can help resolve it.
Here are my recommended label(s): Feature

[pengzhao-intel]

@eric-haibin-lin is there any use case and background?

[eric-haibin-lin]

Any distributed training job with parameter server uses this for gradient aggregation.

[eric-haibin-lin]

@PatricZhao maybe the right tag is MKL instead of MKLDNN?

[bgawrych]

Hi @eric-haibin-lin - I assume that this feature wasn't done yet - can you provide minimum test for this?