[vtk-developers] Added support for std::vector to the wrappers

[David Gobbi]

Hi Todd,

The Python buffer interface is the opposite of marshalling.  Python objects
that support the buffer interface are able to directly access each other's
memory.  The contents of a vtkDataArray wrapped in Python can be directly
accessed by numpy routines that operate on buffers, i.e. the underlying
C/C++/FORTRAN code directly sees the array contents as a raw pointer plus
the data type, dimensions, and strides.  So, no, it isn't used for strings
arrays or object arrays.  And it isn't supported by tuples since tuples
aren't meant as a bulk data container.

I already mentioned in my first post that if people want to be able to
return large vtk::vectors, I can add a wrapped std::vector object and have
the wrappers return that instead of a tuple (and of course the wrapped
vtk::vector would support iteration and indexing in Python).  There already
are vtkVector and vtkTuple types for small fixed-size arrays, and parts of
VTK's API uses them.  These are wrapped directly in Python rather than
converted to tuples.

 - David


On Thu, Sep 6, 2018 at 3:16 AM Todd Martin <nztoddler at yahoo.com> wrote:

> Hi David
>
> I agree that moving large chunks of data in a buffer is more efficient
> than iteration, but is that how arrays are marshaled into tuples in VTK? It
> certainly isn't possible for string/object arrays, since the memory size
> varies for each item.
>
> I can also see that indexing into a list is more efficient, when a few
> specific items are needed; however, I don't think a tuple is the right
> object type for this, if the count is large. If std::array and std::vector
> are already being used for string lists, I believe that vtkArray and
> vtkVector types should be introduced that, optionally, expose a method to
> return a vtkIterator.
>
>
>
> Todd Martin, PhD.
> *Freelance Engineer/Software Architect.*
>
>
>
> On Thursday, September 6, 2018, 2:29:45 PM GMT+12, David Gobbi <
> david.gobbi at gmail.com> wrote:
>
>
> Hi Todd,
>
> VTK already supports the Python buffer interface for moving large amounts
> of data between VTK and Python, people have been using it to shuffle arrays
> between VTK and numpy for many years.  It's much more efficient than any
> iterator-based scheme.
>
> The main goal of wrapping std::vector is convenience: people have already
> been using std::vector for things like string lists, so there was a present
> demand for this feature.  The reason that I stuck with std::vector rather
> than going for a generic iterator-based interface is that it made it easy
> for me to piggy-back on the wrappers' existing support for C arrays.  Very
> little extra weight was added to the wrappers.
>
>  - David
>
>
> On Wed, Sep 5, 2018 at 7:20 PM Andras Lasso <lasso at queensu.ca> wrote:
>
> I find this feature very useful! In application code, we commonly need to
> deal with variable-length arrays. We have been using std::vector accessors
> in C++ and adding vtkStringArray based or GetNumberOf…/GetNth… methods or
> for Python, which has made the API complex and has been very inconvenient
> to use in Python.
>
>
>
> These arrays typically contain up to a few ten elements, therefore, for
> us, convenience is much more important than efficiency.
>
> Most often we return strings and vtkObject* in std::vector, so it would be
> great if vtkObject pointers would be supported, too.
>
>
>
> Andras
>
>
>
>
>
> *From:* vtk-developers <vtk-developers-bounces at public.kitware.com> *On
> Behalf Of *Todd via vtk-developers
> *Sent:* Wednesday, September 5, 2018 8:16 PM
> *To:* David Gobbi <david.gobbi at gmail.com>
> *Cc:* VTK Developers <vtk-developers at vtk.org>
> *Subject:* Re: [vtk-developers] Added support for std::vector to the
> wrappers
>
>
>
> Hi David
>
>
>
> Since std::vector types are unlikely to be small, like coordinate and
> bound arrays, that sounds horribly inefficient; especially when large
> quantities of strings and objects are being marshalled.
>
>
>
> Why not just map the std::vector/std::array to a Python iterator object?
>
>
>
> Todd
>
>
>
> On 6 Sep 2018 4:20 a.m., David Gobbi <david.gobbi at gmail.com> wrote:
>
> Hi Folks,
>
>
>
> Yesterday I merged a patch that allows std::vector<T> to be used from the
> Python wrappers, where T can be 'std::string' or any numeric type from
> 'signed char' to 'float' (including typedefs to those types).
>
>
>
> This means that if you add a C++ method to VTK that takes
> std::vector<std::string>, then in Python you can pass a list or any other
> Python sequence and the wrappers will convert it to a vector for you.
>
>
>
> Method examples that can accept any Python sequence:
>
>     void SetStrings(std::vector<std::string> strings);
>
>     void SetStrings(const std::vector<std::string>& strings);
>
>
>
> Method example that takes a Python mutable sequence (the C++ method can
> write back to the container, since this is pass-by-reference):
>
>     void GetStrings(std::vector<std::string>& strings);
>
>
>
> Method examples that return a vector (which is converted to a Python
> tuple):
>
>     const std::vector<std::string>& GetStrings();
>
>     std::vector<std::string>& GetStrings();
>
>     std::vector<std::string> GetStrings();
>
>
>
> Note that std::vector is not 'wrapped' per se.  Instead, what is going on
> here is that the wrappers are converting to/from Python container types. My
> primary goal was convenience rather than efficiency, we can always add a
> wrapped std::vector type at a later date without sacrificing the
> convenience of automatic conversion.
>
>
>
> Chances are good that I'll be able to add std::vectors of VTK objects
> (smart or dumb pointers) and vtkVariant before VTK 9, but only if someone
> adds methods to VTK that take parameters of those types (as far as I know,
> there aren't any yet).
>
>
>
> See the following VTK gitlab issue for additional information:
>
> https://gitlab.kitware.com/vtk/vtk/issues/17362
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitlab.kitware.com%2Fvtk%2Fvtk%2Fissues%2F17362&data=02%7C01%7Classo%40queensu.ca%7C0e1f3f218b5543a5524608d6138e03b5%7Cd61ecb3b38b142d582c4efb2838b925c%7C1%7C0%7C636717898021517951&sdata=9vohuaiduQLCm6pfGlMxQbXL4hYutf1fKB6SafvQE4c%3D&reserved=0>
>
>
>
> The following is a list of VTK methods that take std::vector that have
> been wrapped due to this change (it is a short list, since people have
> historically avoided using std::vector as part of the VTK public API due to
> the obvious reason that they didn't want to exclude their methods from
> wrapping):
>
>
>
> vtkActor::ProcessSelectorPixelBuffers
>
> vtkAMRInformation::GetNumBlocks
>
> vtkCompositePolyDataMapper2::ProcessSelectorPixelBuffers
>
> vtkMapper::ProcessSelectorPixelBuffers
>
> vtkMultiProcessStream::GetRawData
>
> vtkMultiProcessStream::SetRawData
>
> vtkOpenGLGlyph3DHelper::GlyphRender
>
> vtkOpenGLIndexBufferObject::AppendTriangleIndexBuffer
>
> vtkOpenGLIndexBufferObject::AppendLineIndexBuffer
>
> vtkOpenGLIndexBufferObject::AppendTriangleLineIndexBuffer
>
> vtkOpenGLIndexBufferObject::AppendPointIndexBuffer
>
> vtkOpenGLIndexBufferObject::AppendStripIndexBuffer
>
> vtkOpenGLIndexBufferObject::AppendEdgeFlagIndexBuffer
>
> vtkOpenGLMoleculeMapper::ProcessSelectorPixelBuffers
>
> vtkOpenGLPointGaussianMapper::ProcessSelectorPixelBuffers
>
> vtkOpenGLPolyDataMapper::HandleAppleBug
>
> vtkOpenGLPolyDataMapper::ProcessSelectorPixelBuffers
>
> vtkOpenGLVertexBufferObject::SetShift
>
> vtkOpenGLVertexBufferObject::SetScale
>
> vtkOpenGLVertexBufferObject::GetShift
>
> vtkOpenGLVertexBufferObject::GetScale
>
> vtkOpenGLVertexBufferObject::GetPackedVBO
>
> vtkParallelAMRUtilities::DistributeProcessInformation
>
> vtkProp::ProcessSelectorPixelBuffers
>
> vtkResourceFileLocator::Locate
>
> vtkShadowMapPass::ShadowMapTransforms
>
> vtkShadowMapPass::GetShadowMapTextureUnits
>
> vtkSparseArray::GetUniqueCoordinates
>
> vtkUnicodeString::utf16_str
>
>
>
> Cheers,
>
>  - David
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://public.kitware.com/pipermail/vtk-developers/attachments/20180906/00fd1949/attachment-0001.html>