Fix integer overflows in pylibcudf from_column_view_of_arbitrary

[wence-]

Description

Although the number of rows in a column_view cannot exceed size_type, the number of bytes surely can. So fix that.

This fixes multi-GPU errors we are seeing in rapidsmpf when communicating buffers.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[vyasr]

Other sources of overflow were fixed in #18734. Those fixes conflict partially with this one, but this fix is also necessary since that PR did not fix the chars case, so this PR is still necessary.

[vyasr]

Thanks for fixing up all the bindings for size_of's return type as well, I didn't do that in my PR to get a fix in quickly but I agree that it is necessary.

[davidwendt]

The default empty constructor for cudf::strings_column_view was left out on purpose but I understand the struggle that is Cython.

[wence-]

The default empty constructor for cudf::strings_column_view was left out on purpose but I understand the struggle that is Cython.

Thanks, yes.

[wence-]

Actually, by using cython's cpp_locals directive, I can avoid the need for the nullary ctor...

[wence-]

Looks like I screwed something up :(

[quasiben]

Currently failing with errors like:

  File "interop.pyx", line 129, in pylibcudf.interop._from_arrow_column
    return Column(pyarrow_object)
  File "column.pyx", line 249, in pylibcudf.column.Column.__init__
    self._init(obj, *args, **kwargs)
  File "/opt/conda/envs/test/lib/python3.10/functools.py", line 926, in _method
    return method.__get__(obj, cls)(*args, **kwargs)
  File "column.pyx", line 324, in pylibcudf.column.Column._
    result.col.swap(c_result)
AttributeError: C++ attribute 'col' is not initialized

[Matt711]

At least some of the tests pass if you initialize c_children

diff --git a/python/pylibcudf/pylibcudf/column.pyx b/python/pylibcudf/pylibcudf/column.pyx
index 41d7faad08..c5c056c049 100644
--- a/python/pylibcudf/pylibcudf/column.pyx
+++ b/python/pylibcudf/pylibcudf/column.pyx
@@ -353,7 +353,7 @@ cdef class Column:
 
         # TODO: Check if children can ever change. If not, this could be
         # computed once in the constructor and always be reused.
-        cdef vector[column_view] c_children
+        cdef vector[column_view] c_children = vector[column_view]()
         with gil:
             if self._children is not None:
                 for child in self._children:
@@ -388,7 +388,7 @@ cdef class Column:
         if self._mask is not None:
             null_mask = <bitmask_type*>self._mask.ptr
 
-        cdef vector[mutable_column_view] c_children
+        cdef vector[mutable_column_view] c_children = vector[mutable_column_view]()
         with gil:
             if self._children is not None:
                 for child in self._children:

[vyasr]

Given the nature of the changeset, some bug with the cpp_locals directive seems like the most likely culprit.

[Matt711]

Given the nature of the changeset, some bug with the cpp_locals directive seems like the most likely culprit.

Ok will try reverting ea10cb3

[wence-]

Given the nature of the changeset, some bug with the cpp_locals directive seems like the most likely culprit.

Yeah, thanks all.

[wence-]

/merge