Re: [matplotlib-devel] Kivy backend

[Ryan M. <rm...@gm...>]

On Fri, Mar 13, 2015 at 12:53 PM, Chris Barker <chr...@no...>
wrote:

> On Fri, Mar 13, 2015 at 10:21 AM, Benjamin Root <ben...@ou...> wrote:
>
>>  Probably what I am most interested in from OpenGL is its transforms
>> stack.
>>
>
> OpenGL can't do anything with transforms that you couldn't do in python
> (or C, or Cython). What it can do is push the transform computations to the
> GPU(s) -- making for monstrously faster performance.
>
> This is the "problem" with the current MPL architecture. It does all the
> transforming outside of the back-ends, and assumes that the backends can
> only render in 2-d pixel coordinates.
>
> If we can re-factor to push the transforms to the back-end, most of them
> could use the same generic code, but you'd have the option of the back-end
> providing the transforms, which would buy you a LOT with Open GL, and could
> maybe by you some with, say, wxAgg, as you could put the transforms in
> C/C++ perhaps more efficiently.
>
> Note that with OPenGL in general, its the transforming that buys you
> performance -- when you push brand new data to be rendered, it takes a lot
> of time to push that data to the video card, so drawing the first time
> doesn't buy you much. But if you need to re-render that same data in a
> different view, say zooming in or out, etc, then GL can fly -- if that
> transformation can be done on the GPU.
>
>
Don't overestimate the cost of the data transfer, at least for non-millions
of points. Even years ago, just doing basic opengl plotting (no
sophisticated use of on-GPU memory), was a big win.

The other win for mpl + opengl is giving a real rasterizer and depth buffer
for the mpl3d package, which is hampered by the per-artist z-ordering.

Ryan

-- 
Ryan May