Currently, this assumes that cudnn is installed in
/usr/local/cuda/lib64 and /usr/local/cuda/include, because that's what
my current machine has it installed. This can be cleaned up later.

Double check that we correctly transpose height and width everywhere!

b17 disk is no longer writable.

No longer publicly expose the default constructor.

That could allow someone to do something like this:

CuDevice device1;
CuDevice device2;

We're not sure what this would cause, but it probably wouldn't be good.

Document data formats expected of each member function.

There may be edge cases in the configure script with this remaining.

We still don't look up the version of CUDNN to download for your CUDA
version.

Also make sure that non-CUDA builds still compile. This is achieved by
the cu-cudnn-helper.h file, which forward declares cudnn types, which
fortunately are all pointer types (except for enums, which don't matter
in this case).

…nnv7

Remove support for CUDA versions that don't support CUDNN v7.

Remove 32-bit CUDA mode, since no recent version of CUDA supports it. I
may not have removed all instances of it.

Automatically download CUDNN as part of configure.

Convert assertion to warning, although I am not sure this works, since
if an algorithm fails to be found because of an out-of-memory error, it
is likely to fail at training time for the same reason.

Use cudnnSetTensor4dDescriptor with strides we calculate ourselves
instead.

Change filter type back to NCHW, since it supports more algos.

It supports only the precomputed implicit GEMM implementation of
convolution.

The ConvolutionComputation class does not know whether or not the bias
is optional at initialization time. It will simply avoid using the bias
if it is a nullptr.

Don't use cudnnConvolutionBiasActivationForward.

…ks: johnjosephmorgan@gmail.com