Try to extract Convolution code from cuda-convnet2

[sguada]

According to some benchmarks Caffe was the fastest Convolution until the recent release of cuda-convnet2
https://github.com/soumith/convnet-benchmarks

So maybe someone would like to look into the code and try to extract some code for doing convolutions in Caffe.
https://code.google.com/p/cuda-convnet2/

[bhack]

Could be interesting to see this benchmark with #544 but it is still CPU only.

[kloudkl]

cuda-convnet2 has three major new features relative to cuda-convnet:

1. Improved training times on Kepler-generation Nvidia GPUs (Geforce Titan, K20, K40).
2. Multi-GPU training support implementing data parallelism, model parallelism, and the hybrid approach described in One weird trick for parallelizing convolutional neural networks [1].
3. Less-polished code and incomplete (but improving) documentation.

[1] Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv:1404.5997 [cs.NE]

[kloudkl]

It was said that "Caffe is fastest forward+backward". What is the "banded approach for im2col" in the comment?

[kloudkl]

There have been a lot of interests in running Caffe on multiple GPUs (#194, #301, #423, #519, #547, #630, #653). Alex is keeping ahead of Caffe by implementing the data prallelism, model parallelism, and the hybrid approach. Can something be done to catch up?

[rodrigob]

The memory consumption aspect #852 should also be considered.

[shelhamer]

For parallelism, please discuss at #876. It's certainly planned, but given the freedom in this direction don't hesitate to attempt parallelism after your own fashion for comparison.

[shelhamer]

Closing -- cuDNN and the Caffe layer integration supersedes the custom cuda-convnet2 kernels.