Also in kernel_codegen.py There is a fix then X is strided. The
variable loop_line wasn't checking the stride.

Revert "msvc doesn't understand -Wl,-rpath option"

This reverts commit cc639be5.

This shouldn't be done on the platform, but on the compiler. Anyway,
we won't make GPU op link to BLAS in Pylearn2 for now, so I just
remove this.

msvc doesn't understand -Wl,-rpath option

Added 'on_opt_error=pdb' to fall into pdb on error

Minor fixes

This makes it easier to debug crashes in optimizations.

Also made sample code better PEP8-compliant, and fixed a minor typo in
an error message.

Fixes gh-1353

Add a note in the doc with URL to help link with MKL.

make sparse Arctanh gradient test more stable.

The dostring of the file tensor/basic.py isn't automatically included
in the theano doc as others module as it don't display nicely. So we
do it in a better format manually.

fix gh-1317

Otherwise, it was failing due to numerial problem in some case.

Removed a misleading comment.

Add a benchmark result for the GTX 650 Ti

Add a benchmark result for the GTX 650 Ti and add spaces in the card names so that this file can more easily be found when searching (via grep, Google, etc.) using the real names

The value added to a_i (result of multiplication i_d*a_str[d]) was unsigned int, which broke when strides were negative.
This is not a bug in nvcc 3.0, it is correct behavior according to the C99 standard:

"Otherwise, if the operand that has unsigned integer type has rank greater or equal to the rank of the type of the other operand, then the operand with signed integer type is converted to the type of the operand with unsigned integer type."

Always use the value of rebuild_strict