Revert a recent change. Otherwise some tests break in FAST_COMPILE and we end up with not as much optimized code.

Add printing in the gpu example to help debug report problem (people frequently forget to set floatX=float32.

GpuDnnBatchNorm with 5d inputs

buildbot explicit cuda path

theano.tensor.signal.Pool with 3D support

Adding an AbstractConv3d interface

The new Corr3dMM opts are also used to optimise AbstractConv.
The code is similar to the 3D version in corr_gem.c.

This reuses the implementation of GpuCorr2dMM and its gradient ops.

Add abstract convolution classes, reuse this for 2D and 3D.

fix softmaxgrad dnn opt error fixes #5056

Remove ProfileMode

GpuAdvancedIncSubtensor1 supports mixed dtypes.

Change numpy.any to numpy.all in check_equal_numpy