THTensor_(*) for abs, sqrt and tanh

[szagoruyko]

removing legacy code

[soumith]

Thanks

[fmassa]

There is actually a performance loss for Sqrt and Tanh with this PR.
Current pointwise torch functions are not paralellized, see https://github.com/torch/torch7/blob/master/lib/TH/generic/THTensorMath.c#L1854 and https://github.com/torch/torch7/blob/master/lib/TH/generic/THTensorMath.c#L1802
I agree though that nn functions should maximally rely on torch functions. In this case, shouldn't we add parallelization on pointwise functions in core torch ?
With current framework, I think it would require manually writing repeated code for each pointwise function, as it seems that one can not use preprocessor macros (#pragma omp parfor) inside a macro definition.

[hughperkins]

With current framework, I think it would require manually writing repeated code for each pointwise function, as it seems that one can not use preprocessor macros (#pragma omp parfor) inside a macro definition.

Hmmm, plus one point in favor of Jinja2/Lua templates :-P (Mind you, this can also be argued to be true also for c++ templates too plausibly).

[fmassa]

Well, there are workarounds to allow parallel functions inside a macro.
Here is first thing I tried that works (at least on gcc 4.6.3). fmassa/torch7@c0a6654
There are lots of compiler warnings (maybe because we are calling functions like tanh with float arguments, but it should be tanhf instead), but it seems to work.
Any better solutions that work on C99 ?

[hughperkins]

seems like it is because you are sending a pointer to a function that takes a double, into a function that is expecting a pointer to a function that takes a float. Sent a pull request that seems to remove build warnings (no idea if it runs correctly or not though :-P ) fmassa/torch7#1

[szagoruyko]

we have a fix for all math foo/foof cases, this PR was a preparation for it