ENH: Add support for __cuda_array_interface__

Adapt ITK's PyBuffer approach to inject __cuda_array_interface__ in wrapped
code.
Closes RTKConsortium#10

wrapping/CudaImage.i.in

@@ -0,0 +1,17 @@
+%extend itkCudaImage@CudaImageTypes@{
+    %pythoncode %{
+    @property
+    def __cuda_array_interface__(self):
+        _pixelType = "@PixelType@"
+        _typestr = _get_type_string(_pixelType)
+        return {
+            'shape': (self.GetLargestPossibleRegion().GetSize()[0], self.GetLargestPossibleRegion().GetSize()[1], self.GetLargestPossibleRegion().GetSize()[2]),
+            'data': (int(self.GetCudaDataManager().GetGPUBufferPointer()), False),
+            'typestr': _typestr,
+            'descr': [('', _typestr)],
+            'version': 3,
+            'stream': None,
+            'strides': None
+        }
+    %}
+};

wrapping/CudaImage.i.init

@@ -0,0 +1,37 @@
+%pythoncode %{
+def _get_type_string(itk_Image_type):
+    """Returns the type string of the ITK PixelType as defined in the NumPy array interface."""
+
+    # This is a Mapping from system byte order to typestr byte order.
+    _byteorder_typestr = {
+        "big":">",
+        "little":"<",
+    }
+    import sys
+    _byteorder = _byteorder_typestr[sys.byteorder]
+
+    # This is a Mapping from itk pixel types to typestr.
+    _itk_typestr = {
+        "UC":"|u1",
+        "US":_byteorder + "u2",
+        "UI":_byteorder + "u4",
+        "UL":_byteorder + "u8",
+        "ULL":_byteorder + "u8",
+        "SC":"|i1",
+        "SS":_byteorder + "i2",
+        "SI":_byteorder + "i4",
+        "SL":_byteorder + "i8",
+        "SLL":_byteorder + "i8",
+        "F":_byteorder + "f4",
+        "D":_byteorder + "f8",
+    }
+    import os
+    if os.name == 'nt':
+        _itk_typestr['UL'] = _byteorder + "u4"
+        _itk_typestr['SL'] = _byteorder + "i4"
+
+    try:
+        return _itk_typestr[itk_Image_type]
+    except KeyError as e:
+        raise e
+%}

wrapping/itkCudaImage.wrap

@@ -1,9 +1,15 @@
-itk_wrap_class("itk::CudaImage" POINTER_WITH_CONST_POINTER)
+configure_file("${CMAKE_CURRENT_SOURCE_DIR}/CudaImage.i.init" "${CMAKE_CURRENT_BINARY_DIR}/CudaImage.i" @ONLY)
 
+itk_wrap_class("itk::CudaImage" POINTER_WITH_CONST_POINTER)
   UNIQUE(types "UC;UL;${ITKM_IT};${WRAP_ITK_SCALAR}")
-  foreach(d ${ITK_WRAP_IMAGE_DIMS})
-    foreach(t ${types})
+  foreach(t ${types})
+    set(PixelType ${t})
+    foreach(d ${ITK_WRAP_IMAGE_DIMS})
       itk_wrap_template("${t}${d}" "${ITKT_${t}}, ${d}")
+      set(CudaImageTypes ${t}${d})
+      configure_file(${CMAKE_CURRENT_SOURCE_DIR}/CudaImage.i.in ${CMAKE_CURRENT_BINARY_DIR}/CudaImage.i.temp @ONLY)
+      file(READ ${CMAKE_CURRENT_BINARY_DIR}/CudaImage.i.temp CudaImageInterfaceTemp)
+      file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/CudaImage.i ${CudaImageInterfaceTemp})
     endforeach()
   endforeach()
 
@@ -17,3 +23,21 @@ itk_wrap_class("itk::CudaImage" POINTER_WITH_CONST_POINTER)
   endforeach()
 
 itk_end_wrap_class()
+
+# Add library files to be included at a submodule level and copy them into
+# ITK's wrapping typedef directory.
+# Another approach is to add CudaImage.i to the WRAPPER_SWIG_LIBRARY_FILES list
+# but then the %pythoncode from CudaImage.i.init gets only included in
+# itkCudaDataManagerPython.py even if the WRAPPER_SUBMODULE_ORDER is set.
+# Prefer using ITK_WRAP_PYTHON_SWIG_EXT to make sure the block is included in
+# the right file exclusively.
+set(ITK_WRAP_PYTHON_SWIG_EXT
+  "%include CudaImage.i\n${ITK_WRAP_PYTHON_SWIG_EXT}")
+
+file(COPY "${CMAKE_CURRENT_BINARY_DIR}/CudaImage.i"
+  DESTINATION "${WRAPPER_MASTER_INDEX_OUTPUT_DIR}")
+
+# Make sure to rebuild the python file when CudaImage.i is modified.
+# Touching CudaImage.i directly does not force a rebuild because it is just
+# appended to the ITK_WRAP_PYTHON_SWIG_EXT variable
+file(TOUCH ${WRAPPER_MASTER_INDEX_OUTPUT_DIR}/itkCudaImage.i)