Implementation convolutionTranspose operator for mkl-dnn (#12866)
Summary:
the speed-up of a single operation is up to 2-3X on BDW.
This PR depend on https://github.com/pytorch/pytorch/pull/14308
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12866
Differential Revision: D13936110
Pulled By: ezyang
fbshipit-source-id: 34e3c2ca982a41e8bf556e2aa0477c999fc939d3