Remove the `double round(double)` in pymath.h for Win32 only by defining
`HAVE_ROUND` via the `-D` switch of `nvcc`.
The `double round(double x)` function is present in CUDA
6.5.13/6.0.37/5.5.22/5.5.11/4.1.28/4.0.17, so I think it's safe to remove
the one in `pymath.h`.

Moved `#include <algorithm>` outside the `#if PY_MAJOR_VERSION >= 3` block.

The generated mod.cu contains a number of std::min and std::max. Including <algorithm> will remove the not-found errors from these two functions.

Fix errors in SparseBlockOuterSS c_code.

Mixed stuff: pep8, comments, doc, don't enable GpuConv*3D op by default at that module import.

[opt warning] fix gpu reduce opt warning.

This allow to move more computation to the gpu.

Float64

Add an option to desactivate conv2d_fft at the op level

3dfftconv

Don't make the kernels debuggable by default, but add those arguments to the 'nvcc-safe' list.

Conv gemm non-square kernel support

Extend documentation (plus small fix)

Cleanup CUDA convolution tests

Add a blocksparse multiplication implementation