Make matplotlib running on GPU/CUDA

Greetings to everyone,

I’m Francesco Faccenda and I’m developing a sfotware in python where
I need to show some charts, so I chose matplotlib as a plot library. Anyway,
since lots of data are shown on matplotlip stages (thousands of curves with
hundreds of points each), the charts become very heavy, interactive events and
functions start lagging and users experience becomes unacceptable.

But there’s a good news, I have a nice GPU available (an NVIDIA Tesla
K40c), so I’d like to know if there is a way to make matplotlib run on it, or
maybe wrap it on some GPU/CUDA wrapper and make it run smoothly.

Do you have any suggestion how to do it? I never went through matplotlib
source code before and I don’t know if this feature has already been scheduled
for future releases or it’s something completely new.

I would really appreciate some suggestions on this subject.

Thank you in advance.

Best regards,

Francesco

I tihnk you want VisPy:

https://vispy.readthedocs.io/en/latest/

It's a plotting package with a kinda like matplotlib API, built on OpenGL.

Unfortunately, it doesn't look like it's been updated in a while -- from
teh docs. But the gitHub project is active:

https://github.com/vispy/vispy

So maybe it's only the docs that haven't been updated!

But probably a much better option than trying to shoehorn GPU rendering
into MPL.

The problem is that while MPL was designed to be "backend" independent --
so it is "easy" to plug in an alternative renderer, the rendering model is
not really well suited to GPU rendering -- it would take a lot of
refactoring to really be able to take advantage of the graphics card.

-CHB

···

On Tue, Sep 12, 2017 at 8:47 AM, Francesco Faccenda <f.faccenda86@...149...> wrote:

But there’s a good news, I have a nice GPU available (an NVIDIA Tesla
K40c), so I’d like to know if there is a way to make matplotlib run on it,
or maybe wrap it on some GPU/CUDA wrapper and make it run smoothly.

--

Christopher Barker, Ph.D.
Oceanographer

Emergency Response Division
NOAA/NOS/OR&R (206) 526-6959 voice
7600 Sand Point Way NE (206) 526-6329 fax
Seattle, WA 98115 (206) 526-6317 main reception

Chris.Barker@...236...

Thank you Chris for your reply.
I have to admit I already stumbled on VisPy while doing my research on the web. Still, I’ve got a lot of code already working with matplotlib. Indeed, not only I plot data with it, but i manage a lot of mpl events to provide the users usefool tools, like lines picking, tooltips, lines copy/paste, square selectors for multiple selections, context menu and so on. Moreover, I got matplotlib embedded on wxpython as well. While at the beginning few lines were managed and noone complained, now that big amout of data has to be displayed, the non-GPU core of the library is starting to show its limits.

Since matplotlib is a reference library for this kind of applications, I thought it deserved an update in this direction. If anyone is willing to do so, I’m available to discuss possible solutions and also provide any help I can give.

Best regards,

Francesco Faccenda

···

2017-09-13 0:46 GMT+02:00 Chris Barker <chris.barker@…236…>:

On Tue, Sep 12, 2017 at 8:47 AM, Francesco Faccenda <f.faccenda86@…149…> wrote:

But there’s a good news, I have a nice GPU available (an NVIDIA Tesla
K40c), so I’d like to know if there is a way to make matplotlib run on it, or
maybe wrap it on some GPU/CUDA wrapper and make it run smoothly.

I tihnk you want VisPy:

https://vispy.readthedocs.io/en/latest/

It’s a plotting package with a kinda like matplotlib API, built on OpenGL.

Unfortunately, it doesn’t look like it’s been updated in a while – from teh docs. But the gitHub project is active:

https://github.com/vispy/vispy

So maybe it’s only the docs that haven’t been updated!

But probably a much better option than trying to shoehorn GPU rendering into MPL.

The problem is that while MPL was designed to be “backend” independent – so it is “easy” to plug in an alternative renderer, the rendering model is not really well suited to GPU rendering – it would take a lot of refactoring to really be able to take advantage of the graphics card.

-CHB

Christopher Barker, Ph.D.
Oceanographer

Emergency Response Division
NOAA/NOS/OR&R (206) 526-6959 voice
7600 Sand Point Way NE (206) 526-6329 fax
Seattle, WA 98115 (206) 526-6317 main reception

Chris.Barker@…236…

I have to admit I already stumbled on VisPy while doing my research on the
web. Still, I've got a lot of code already working with *matplotlib*.
Indeed, not only I plot data with it, but i manage a lot of *mpl events*
to provide the users usefool tools, like lines picking, tooltips, lines
copy/paste, square selectors for multiple selections, context menu and so
on. Moreover, I got matplotlib *embedded *on *wxpython *as well. While at
the beginning few lines were managed and noone complained, now that big
amout of data has to be displayed, the non-GPU core of the library is
starting to show its limits.

Since matplotlib is a reference library for this kind of applications, I
thought it deserved an update in this direction.

Well, As I understand it, VisPY made some effort to be compatible with the
MPL API -- but that is going to depend on how much you use the lower-level
parts f the API -- things like the transform, etc. to take advantage of GPU
rendering, all the transforms, etc needs to be pushed to the GPU, so the
architecture (and API) needs to be quite different.

If anyone is willing to do so, I'm available to discuss possible solutions
and also provide any help I can give.

As Ben pointed out, and I was trying to make clear -- it really isn't a
matter of "just" writing an OpenGL backend -- there really needs to be a
major restructuring.

And VisPy is pretty much that project. Whether it is feature complete,
robust and maintained enough for your use-cases, I have no idea, but even
if not -- you'll probably be better off contributing to that effort than
starting all over with trying to make an GPU_based OPenGL back-end.

However -- maybe there is another option:

Taking full advantage of GPUs does require a full restructuring, but maybe
there are other ways to get better performance -- for instance, optimizing
the transform code, etc:

Using the GPU with PyCuda or [what the heck is the name of the more general
GPU-using lib??]

using numba

Maybe there is still room for Cython, etc....

In short, profiling MPL carefully, to see where the performance bottlenecks
are:

With modern hardware, actually rendering stuff is no longer the slow part
of visualization. Rather, it's pushing data to the renderer, transforming
data etc.

This is why to take advantage of the GPU, you need to do the
transformations, etc on the GPU -- which the MPL architecture doesn't make
easy by dropping in a new back-end.

Which is why VisPy built a new architecture from the bottom up.

-CHB

···

On Wed, Sep 13, 2017 at 12:31 AM, Francesco Faccenda <f.faccenda86@...149... > wrote:

--

Christopher Barker, Ph.D.
Oceanographer

Emergency Response Division
NOAA/NOS/OR&R (206) 526-6959 voice
7600 Sand Point Way NE (206) 526-6329 fax
Seattle, WA 98115 (206) 526-6317 main reception

Chris.Barker@...236...

Depending on the exact use case you can get pretty good mileage out of blitting (See http://matplotlib.org/api/animation_api.html#funcanimation for an explanation or how it is used in the widgets module).

The best way to make things faster is to just do less work :slight_smile:

Tom

···

On Wed, Sep 13, 2017 at 12:31 AM, Francesco Faccenda <f.faccenda86@…149…> wrote:

I have to admit I already stumbled on VisPy while doing my research on the web. Still, I’ve got a lot of code already working with matplotlib. Indeed, not only I plot data with it, but i manage a lot of mpl events to provide the users usefool tools, like lines picking, tooltips, lines copy/paste, square selectors for multiple selections, context menu and so on. Moreover, I got matplotlib embedded on wxpython as well. While at the beginning few lines were managed and noone complained, now that big amout of data has to be displayed, the non-GPU core of the library is starting to show its limits.

Since matplotlib is a reference library for this kind of applications, I thought it deserved an update in this direction.

Well, As I understand it, VisPY made some effort to be compatible with the MPL API – but that is going to depend on how much you use the lower-level parts f the API – things like the transform, etc. to take advantage of GPU rendering, all the transforms, etc needs to be pushed to the GPU, so the architecture (and API) needs to be quite different.

If anyone is willing to do so, I’m available to discuss possible solutions and also provide any help I can give.

As Ben pointed out, and I was trying to make clear – it really isn’t a matter of “just” writing an OpenGL backend – there really needs to be a major restructuring.

And VisPy is pretty much that project. Whether it is feature complete, robust and maintained enough for your use-cases, I have no idea, but even if not – you’ll probably be better off contributing to that effort than starting all over with trying to make an GPU_based OPenGL back-end.

However – maybe there is another option:

Taking full advantage of GPUs does require a full restructuring, but maybe there are other ways to get better performance – for instance, optimizing the transform code, etc:

Using the GPU with PyCuda or [what the heck is the name of the more general GPU-using lib??]

using numba

Maybe there is still room for Cython, etc…

In short, profiling MPL carefully, to see where the performance bottlenecks are:

With modern hardware, actually rendering stuff is no longer the slow part of visualization. Rather, it’s pushing data to the renderer, transforming data etc.

This is why to take advantage of the GPU, you need to do the transformations, etc on the GPU – which the MPL architecture doesn’t make easy by dropping in a new back-end.

Which is why VisPy built a new architecture from the bottom up.

-CHB

Christopher Barker, Ph.D.
Oceanographer

Emergency Response Division
NOAA/NOS/OR&R (206) 526-6959 voice
7600 Sand Point Way NE (206) 526-6329 fax
Seattle, WA 98115 (206) 526-6317 main reception

Chris.Barker@…236…