Tensor Type has now a sparse grad attribute which can be set or unset to determine if sparse gradient is to be enabled or not. Note if needed, we can provide getters and setters to set the attributes to provide more clean interfaces to application. Right now application can access this attribute and set it before calculating gradient

Issue #1164: Fixed failing test_pycuda_example.

Removed unnecessary #includes from generated cuda code.

Denormal

fix missing space in error message

Fixed compile errors under llvm-gcc-4.2 (OSX 10.8).

Ready to merge: get rid of dangerous "TypeError = not constant" mechanism

Gpu sum crash fix

fix race condition

added in the 2 months since I originally submitted the pull request

Popen windows

Always define the name attribute to a default value.

even though it is documented as being 0d

as being constant