* - MKLDNN would choose the algorithm which will potentially give best performance based on
- convolution dimensions number of logical processors available.

- (For auto-dispatching to work as intended,
- use the same thread affinity settings when creating the convolution as when executing the convolution.)
- The relationship between convolution sizes and the best performing algorithm is empirically based on performance observations

* bump mkldnn version to V0.18-rc

* Revert "- MKLDNN would choose the algorithm which will potentially give best performance based on"

This reverts commit 904beb8ad8d4e829fbae5f38a803ea80a72b3ffd.

* Update mkl-dnn patch for soversion removal.