COMPAT: use tkagg backend on PyPy #9356

mattip · 2017-10-11T01:19:22Z

PR Summary

the tkagg::blitfunction uses the id built-in to obtain memory adresses. This is not compatible with PyPy. I have written a proof-of-concept alternative blit function that uses a pure-python _tkagg backend via CFFI. Since CPython prefers a c-extension (_tkagg.so) over a python file (_tkagg.py) and PyPy reverses this, the two implementations can live side by side.

It is working but not yet complete:

~~on the matplotlib side I should probably only conditionally build the _tkagg c-extension on CPython~~
~~on the PyPy side the serialization/deserialization of a numpy ndarray should not make a copy of the data~~
~~updating only a bbox of data not supported yet~~
~~documentation~~

but I wanted to see if this approach is acceptable before continuing. Any thoughts?

Edit: now point three (bbox handling) is complete
Edit: (5.Nov.17) now points one and two are done, and documentation does not change

anntzer · 2017-10-11T02:49:57Z

Looks like the only problematic use of id() is to get the underlying address of the ndarray object (the PyObject, not the memory buffer), right? I would suggest instead just passing a memoryview object, which contains all the relevant information even in Py2 (https://docs.python.org/2/library/stdtypes.html#memoryview), and adapting the extension code to accept such an input (it's all private anyways).

Actually now that I look at it it's not clear whether tk.call can pass memoryviews, but you can also just pass in the address of the underlying buffer and the size and strides of the array (and fix the extension code accordingly). The address of the underlying buffer is accessible at t.__array_interface__["data"] (https://docs.scipy.org/doc/numpy-1.13.0/reference/arrays.interface.html#__array_interface__).

If possible, I'd prefer avoid too much pypy specific code, because I fear it's quite likely to bitrot...

mattip · 2017-10-16T18:41:28Z

@anntzer thanks for picking this up.

I tried to modify the id() calls as you said, but have not been able to get load_tkinter_funcs fron the _tkagg.cpp c-extension working on PyPy, so in the mean time I fleshed out this approach to see if it is functional.

…thon3

mattip · 2017-11-05T19:39:30Z

I actually needed to change very little to get this working, in the end I:

modified the private calls to pass "ptr height width" instead of id(array)
added another case to the code on non-win32 that searches for the shared object that works with PyPy's cffi module tk._tkinter.tklib_cffi.__file__

On PyPy unfortunately there seems to be an issue with opening the "configure subplot" subfigure on a basic plot, the mouse events are not reliably reaching the event handlers in lib/matplotlib/widgets.py, maybe a threading issue or competing event loops?

anntzer · 2017-11-05T19:58:51Z

lib/matplotlib/backends/tkagg.py

+            _tkagg.tkinit(tk.interpaddr(), 1)
+        else:
+            # very old python?
+            _tkagg.tkinit(tk, 0)


interpaddr was added in 1998(! python/cpython@9d1b7ae) so I cannot actually imagine a case where this is triggered today (but that can go to another PR of course)

mattip · 2017-11-07T18:07:33Z

@tacaswell I got a mail reviewing the changes to setupext.py but cannot see it here, I removed the reference to py_converters.cpp since it is no longer needed. It was used to convert the python-level id(ndarray) to a C-level ndarray. Now I changed the code to pass in ptr height width instead, so the conversion does not require any knowlege about ndarrays

mattip · 2017-11-07T18:08:20Z

How do I fix the codecov/patch check failure?

dopplershift · 2017-11-07T18:49:58Z

Don't worry about the codecov check there. The Tk GUI code, unfortunately, has no automated tests for CI to run.

anntzer · 2017-11-07T18:56:03Z

We tend to ignore that check to be honest...

mattip · 2017-11-23T21:46:34Z

ping

anntzer · 2017-11-23T21:55:30Z

src/_tkagg.cpp

 #endif
+# define BBOX_FORMAT "%f %f %f %f"


extra space?

fixed in cc7b73e

anntzer · 2017-11-23T21:56:19Z

src/_tkagg.cpp

        deststride = 4 * destwidth;

-        destbuffer = new agg::int8u[deststride * destheight];
+        destbuffer = new uint8_t[deststride * destheight];


any reason for the change?

now it no longer uses #include py_converters.h which supplied the agg:int8u

anntzer · 2017-11-23T22:00:50Z

src/_tkagg.cpp

-            return TCL_ERROR;
-        }
-
+    if (sscanf(argv[4], BBOX_FORMAT, &x1, &x2, &y1, &y2) == 4) {


should probably error out cleanly if an invalid string is passed?

yesish, these are basically "private" functions parsing a call from python, but I will add more checking code

updated in cc7b73e

anntzer · 2017-11-23T22:17:19Z

src/_tkagg.cpp

        has_bbox = true;
    }
-    else {
+    else if (bbox_parse == 1 && x1 == 0){


anntzer · 2017-11-24T00:12:06Z

Can you rebase on master (to try to make appveyor pass... perhaps)?

anntzer · 2017-11-24T07:49:35Z

src/_tkagg.cpp

-        desty = hdata - y2;
-        destwidth = x2 - x1;
-        destheight = y2 - y1;
+        destx = (int)x1;


Shouldn't BBOX_FORMAT just be 4 integers and just use integers all along? (not sure if tk would have the bad idea to always cast things to float)

Edit: Actually I guess not, what should really happen is that we should explicitly cast on the caller's side but this can just go as it is for now...

The call sequence is backend.blit(image, bbox) -> tkagg.py:blit(image, bbox) -> PyAggImagePhoto(as_string(image, bbox)) so we cannot change the root backend.blit call. AFAICT the choice is:

convert early in tkagg.py:blit

convert later in PyAggImagePhoto

I usually prefer converting as late as possible to postpone rounding and allow future refactoring with floats to be as painless as possible, but will flow with whatever is matplotlib best practices

I think we can leave it as it is for now, we were using floats to start with and we can change that later.

anntzer

assuming appveyor passes

Tested locally on cpython, not actually tested myself on pypy...

mattip · 2017-11-24T10:21:32Z

thanks, I tested on pypy HEAD and tkagg now works. However the way matplotlib uses weakref.ref in callbacks exposes a bug in the JIT, am working on distilling it so I can report and get it fixes

mattip · 2017-11-28T22:08:52Z

just to be clear, the weakref issue is not a matplotlib issue, rather a PyPy issue that matplotlib exposes

mattip · 2017-12-02T15:47:39Z

after #9899 was merged, this is all that is needed for the next release of PyPy to work with the TkAgg backend

embedding_in_tk2_sgskip is essentially a slightly simplified duplicate of embedding_in_tk_sgskip.

tacaswell · 2017-12-11T19:28:15Z

src/_tkagg.cpp

        deststride = 4 * destwidth;

-        destbuffer = new agg::int8u[deststride * destheight];
+        destbuffer = new uint8_t[deststride * destheight];


Why change away from the agg types here?

Previously there was a conversion from a PyArrayObject * (in other words ndarray) to a data pointer, width, height, etc via py_converters. Now I pass that information directly from python. I no longer need the "py_converters.cpp" and ""py_converters.h" since there is no use of ndarray in the c++ code. Should I leave the include for this one type definition?

I am concerned about the corner case where unint8_t and agg:int8u are different underlying types. Not sure how likely that actually is, hence leaning towards being conservative about changing it.

I wonder what different types you think these can be.
I would also bet that there must be parts of the code where we assume that this is interconvertible with np.uint8, or that this fits in a single byte... doesn't leave you a lot of options!

mattip · 2017-12-11T20:15:09Z

I never mentioned that what I thought was a PyPy weakref or JIT issue was actually solved by PR #9899 so this PR is all that is needed for TkAgg and PyPy

mattip · 2017-12-13T06:01:18Z

what's up with docs build? It seems to have timed out with no connection to my pull request

@tacaswell converted back to agg::int8u

mattip · 2017-12-19T21:04:07Z

what's up with docs build? It seems to have timed out with no connection to my pull request

@tacaswell converted back to agg::int8u

edit: docs build seems to have passed

jklymak · 2017-12-19T21:12:33Z

Sometimes the builds need to be restarted manually because the host has a problem...

mattip · 2018-01-11T20:47:18Z

can I get this marked 2.2-critical?

anntzer · 2018-01-11T23:58:51Z

"Making PyPy work" sounds like a good thing to have...

anntzer · 2018-01-12T00:06:00Z

By the way would be nice to add a whatsnew entry (including info about what versions of PyPy are supposed to work).

mattip · 2018-01-12T11:22:25Z

Not sure 270f05d is the proper way to do it, let me know if I should have done it differently

tacaswell · 2018-01-12T22:55:50Z

Perfect @mattip

tacaswell · 2018-01-12T22:57:47Z

Thanks @mattip !

tacaswell added this to the 2.2 (next feature release) milestone Oct 11, 2017

tacaswell added the GUI: tk label Oct 11, 2017

mattip added 4 commits November 5, 2017 21:24

simple plot works

b4e2b8a

flesh out bbox handling, fix mpl_PyFile_Dup for PyPy which is like py…

68492fb

…thon3

change interface in PyAggImagePhoto to avoid id()

8e80a20

search for cffi c-extension module to find c-functions

512e4d8

anntzer reviewed Nov 5, 2017

View reviewed changes

pep8

4c5dfaa

anntzer reviewed Nov 23, 2017

View reviewed changes

fixes from review

c2a7143

silence warnings, fix build on old MSC compilers

09576fa

anntzer reviewed Nov 24, 2017

View reviewed changes

anntzer approved these changes Nov 24, 2017

View reviewed changes

mattip mentioned this pull request Dec 1, 2017

make SubplotTool into a modal dialog, keep ref to SubplotTool #9899

Merged

6 tasks

Cleanup embedding_in_tk, remove embedding_in_tk2.

596345e

embedding_in_tk2_sgskip is essentially a slightly simplified duplicate of embedding_in_tk_sgskip.

tacaswell mentioned this pull request Dec 5, 2017

Matplotlib 2 is extremely slow after plt.show() #8129

Closed

mattip mentioned this pull request Dec 8, 2017

Python crashes when closing figures using TkAgg on Mac OS #9856

Closed

Merge branch 'tkagg-anim' into tkagg-cffi

1712678

mattip mentioned this pull request Dec 8, 2017

BUG: clear events before destroying windows in tkagg #9956

Merged

6 tasks

tacaswell reviewed Dec 11, 2017

View reviewed changes

use agg:int8u not uint8_t

b40923d

anntzer mentioned this pull request Dec 15, 2017

FigureCanvasTkAgg - "buffer is of wrong type" error during blit #9799

Closed

Merge branch 'master' into tkagg-cffi

43d9754

anntzer added the Release critical For bugs that make the library unusable (segfaults, incorrect plots, etc) and major regressions. label Jan 11, 2018

add whatsnew entry

270f05d

tacaswell merged commit 3bb328f into matplotlib:master Jan 12, 2018

QuLogic modified the milestones: needs sorting, v2.2.0 Feb 12, 2018

tacaswell mentioned this pull request Feb 15, 2018

TclError in TkAgg with version 2.2.0rc1 #10479

Closed

Uh oh!

COMPAT: use tkagg backend on PyPy #9356

COMPAT: use tkagg backend on PyPy #9356

Uh oh!

Conversation

mattip commented Oct 11, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Summary

Uh oh!

anntzer commented Oct 11, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mattip commented Oct 16, 2017

Uh oh!

mattip commented Nov 5, 2017 • edited by dopplershift Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mattip commented Nov 7, 2017

Uh oh!

mattip commented Nov 7, 2017

Uh oh!

dopplershift commented Nov 7, 2017

Uh oh!

anntzer commented Nov 7, 2017

Uh oh!

mattip commented Nov 23, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anntzer commented Nov 24, 2017

Uh oh!

anntzer Nov 24, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anntzer left a comment

Choose a reason for hiding this comment

Uh oh!

mattip commented Nov 24, 2017

Uh oh!

mattip commented Nov 28, 2017

Uh oh!

mattip commented Dec 2, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mattip commented Dec 11, 2017

Uh oh!

mattip commented Dec 13, 2017

Uh oh!

mattip commented Dec 19, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jklymak commented Dec 19, 2017

mattip commented Oct 11, 2017 •

edited

Loading

anntzer commented Oct 11, 2017 •

edited

Loading

mattip commented Nov 5, 2017 •

edited by dopplershift

Loading

anntzer Nov 24, 2017 •

edited

Loading

mattip commented Dec 19, 2017 •

edited

Loading