GPU implementation of L_p feature pooling

[wickedfoo]

This diff contains the GPU kernels and calling code from the fbcunn library for FeatureLPPooling that we are contributing to cunn itself, converted to compile with modern cutorch / cunn with half, float and double support.

https://github.com/facebook/fbcunn/blob/master/src/FeatureLPPooling.cu https://github.com/facebook/fbcunn/blob/master/src/FeatureLPPoolingHost.cpp

@soumith said he'd add the rest from here.

[wickedfoo]

Missing some updates, one min

[wickedfoo]

ok updated with the header too, and float -> accreal

[wickedfoo]

Please advise if there's anything else that needs to be added on the cunn side. There was no CPU implementation, only this in fbcunn for the module:

https://github.com/facebook/fbcunn/blob/master/fbcunn/FeatureLPPooling.lua

[wickedfoo]

I'm going to write a simple CPU version of this for torch/nn, that shouldn't be too hard.

[soumith]

i think we can make do with the Lua version, dont waste your time.

[wickedfoo]

There's no implementation of the algorithm in Lua either, just the module file that forwards to the cunn version for CudaTensors. Is that good enough?

[soumith]

oh i see. yea it needs a cpu or just pure lua script implementation

[wickedfoo]

I added a CPU implementation in https://github.com/torch/nn/pull/1259

I'll update this cunn pull request with a test of the GPU code that compares against the CPU implementation

[wickedfoo]

This diff has been updated with a test, bugs fixed and it works, compared against the CPU version. This is good to go, as is https://github.com/torch/nn/pull/1259.

[wickedfoo]

ping... can this be merged?