I'm not sure what this test is expecting.  It seems to test that the gpu will magically work and optimize the graph even though we disable InputsToGpuOptimizer (which is obviously not the case).  I wonder why it passed before.  I wonder if this is the right fix.  I also wonder what is it testing exactly.

In any case it seems to pass now.

Scan push out sum of dot

Convert SetSubtensor to IncSubtensor when possible

This did come up in an implementation of adagrad at some point and will be much faster on GPU.

Do not move the GpuSubtensor to the GPU when this do not make sence

[BUG] Fix bug in {inc,set}_subtensor in advanced1 case

Join opt. Remove empty constant as inputs and merge MakeVector inputs.

do the join change for gh-2328

fix gh-2327

Remove useless warning and make the graph more optimized in old supported cases.

When we do this on an inputs/shared var and that this is the only use
of it in the graph, we should only transfer to the GPU the small
computation.

Reported by Stanislas Lauly

Document nvcc.fastmath and fix cxx description.

Tests

Fix problems with CDataType in DebugMode