Add more convenience math functions

[kloudkl]

To fulfill the requests of the issue #146, the following math functions are added and tested: __builtin_popcount* based fast hamming distance, sum of absolute values, element wise sign and fabs, and non-in-place scaling.

Note the cherry pick from the boost-eigen branch. The commit was authored by Rowland Depp. I think it is time to merge the branch into either dev or master.

[sguada]

This should be defined someplace else, probably Makefile.config

[kloudkl]

Commit 4493ab5 (Remove mkl_alternate and move the define func macros to mkl_functions)
deleted the define.

[Yangqing]

Question - why are all files showing as "all adds" instead of modifications? This makes reviewing quite hard.

[kloudkl]

The latest commit 4493ab5 (Remove mkl_alternate and move the define func macros to mkl_functions) avoided the confusion of adding mkl_alternate.hpp. Please make further comments.

[shelhamer]

@kloudkl @Yangqing boost-eigen should most likely be merged into dev before this to make the history more clear and avoid the confusion of the copy-paste cherry picking. I'll look into doing it in such a way that osx 10.9 developers will have a workaround to compile.

[kloudkl]

As pointed out in my previous comment, commit 4493ab5 deleted all the copied contents from the boost-eigen branch to make this PR independent of other issues or PRs. So there is no more urgent need to merge that branch before this.

That being said, the long lasting feature branch boost-eigen has seen no contributing activities for a week and is obviously stabilized.

[shelhamer]

@kloudkl, I did not miss your comment, but I could have explained better. I want boost-eigen in dev to make ongoing work compatible with the integration and avoid further conflicts in the future.

While your c0ceec1 and 4493ab5 manually accomplished this, it would have been unneeded if boost-eigen were already in. The manual code transplant confuses the history in my view, so ideally it could be rebased away, which would be easier if boost-eigen were merged to dev first. I will try this merge with an escape hatch for osx 10.9 to make development more convenient for you, us, and all active developers.

[sergeyk]

@shelhamer What's your ETA on what you're trying to do? Should we just merge this baby in? I feel bad about my Ubuntu-build-breaking HDF5 PR, and there's a commit here that fixes it.

[shelhamer]

This is held on the boost-eigen merge right now. Sorry for the wait @kloudkl – this is a nice contribution. I'll write once boost-eigen is in so you can rebase and we can merge.

[kloudkl]

Rebased and tests passed.

[shelhamer]

Thanks for the further math functions, and comment clarifying the history of the unary function macros. This might cause a slight conflict in the merge of boost-eigen, but that already has its problems and wasn't worth waiting for.

[shelhamer]

@kloudkl this doesn't actually build. The break is at c9d9056 according to git bisect, so it seems to be a problem with the gpu macro.

make: *** Waiting for unfinished jobs....
./include/caffe/util/math_functions.hpp(168): error: expected an identifier

[shelhamer]

@kloudkl I removed this PR from dev given the aforementioned build error. Please open a new PR once the problem is fixed. Thanks again for the contributed functions–looking forward to merging the fixed version.

Note that your demos #141 might conflict with this branch by the implementation of the hamming distance there. Since it is a nice demo, we would like to merge it soon, but please make sure that both are clean merges.

[kloudkl]

Thanks for your review!

I am very sorry for not testing again after rebasing. The new PR that resolves both of your concerns is #201.