Compile with MKL in conda-build

[houseroad]

Problem:
Without -DBLAS=MKL, conda-build won't include MKL library into Caffe2 build. And the BLAS performance is bad on CPU.

Solution:
Explicitly add the flag. Add mkl and mkl-include as dependencies.

@ezyang @Yangqing

[facebook-github-bot]

@houseroad has updated the pull request.

[ezyang]

LGTM. (in chat, I mentioned that it was surprising that mkl-include was needed for the build, since PyTorch doesn't seem to need it. Shrug.)

[houseroad]

Let me check this with Pytorch team.

[facebook-github-bot]

@houseroad has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[Yangqing]

Chatted offline and it's because PyTorch uses explicit BLAS interface only, while C2 MKL mode also uses MKLDNN and a bunch of others. In any case, thanks for the fix! :)

[facebook-github-bot]

@Yangqing has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[uddmorningsun]

@ezyang Any ideas to support caffe2 GPU version conda-build recipe ? Thanks !

[ezyang]

It's supported; at least, I have successfully used an edited version of the conda recipe to build caffe2 with GPU support.