Backends cleanup #8772

anntzer · 2017-06-18T06:21:12Z

This is the first part of #8771, split out for simpler review.

general cleanup to the backends code
unification of blitting and nonblitting code paths in Qt5Agg's paintEvent, similarly to how it is implemented by Gtk3Cairo: a default repaint is simply equivalent to a blitting of the whole canvas.
removal of the unused Gtk3 _needs_redraw attribute (documented as "doesnt work").

tacaswell · 2017-06-19T03:38:53Z

lib/matplotlib/backends/backend_qt4agg.py

-    def __init__(self, figure):
-        if DEBUG:
-            print('FigureCanvasQtAgg: ', figure)
-        FigureCanvasQT.__init__(self, figure)


I think this needs to be here or the MI does not work correctly (as pyqt5 made them cooperative, but they are not int pyqt4)

This is further handled by the qtcairo commit, but in fact with the new MRO (specifically, due to FigureCanvasQT going after FigureCanvasAgg), this is not needed anymore, because the MRO FigureCanvasQT(4)Agg is

matplotlib.backends.backend_qt4agg.FigureCanvasQTAgg matplotlib.backends.backend_qt5agg.FigureCanvasQTAggBase matplotlib.backends.backend_agg.FigureCanvasAgg matplotlib.backends.backend_qt4.FigureCanvasQT matplotlib.backends.backend_qt5.FigureCanvasQT PyQt4.QtGui.QWidget PyQt4.QtCore.QObject sip.wrapper PyQt4.QtGui.QPaintDevice sip.simplewrapper matplotlib.backend_bases.FigureCanvasBase builtins.object

and FigureCanvasAgg.__init__ is the same as FigureCanvasBase.__init__, which does not forward its arguments to the super class, so QWidget never sees the figure kwarg.

Or you can just test by yourself that it works :)

Ah, I was lost in the classes here. It looks like the calls out to the PyQt base classes happen in the FigureCanvasQt.__init__ not here.

🐑

tacaswell · 2017-06-19T03:49:09Z

lib/matplotlib/backends/backend_qt4agg.py

-
-class FigureCanvasQTAgg(FigureCanvasQTAggBase,
-                        FigureCanvasQT, FigureCanvasAgg):
+class FigureCanvasQTAgg(FigureCanvasQTAggBase, FigureCanvasQT):


I think having the base class come in as the _ version of this is clearer (I had to think a bit for why this should work).

Actually this MRO chain will get even further simplified in #8771 (to class FigureCanvasQTAgg(FigureCanvasAgg, FigureCanvasQT), which seems as simple as one can do :-)), sorry for modifying the same thing twice.

I am very much 👍 on simplifying the MRO.

tacaswell · 2017-06-19T03:52:29Z

lib/matplotlib/backends/backend_qt5agg.py

-            # into argb format and is in a 4 byte unsigned int.  Little endian
-            # system is LSB first and expects the bytes in reverse order
-            # (bgra).
-            if QtCore.QSysInfo.ByteOrder == QtCore.QSysInfo.LittleEndian:


Why drop the little endian support?

I'm just reusing the blitting codepath for both cases, so either region.to_string_argb (which is a different method from renderer.tostring_argb...) was already doing the right thing wrt. endianness, or there has been a bug for a while on one of the two endiannesses that no one ever noticed.

fair. There are not many little-endian systems left in practice (and they tend to be big main frame type systems these days).

tacaswell · 2017-06-19T03:56:35Z

lib/matplotlib/backends/backend_qt5agg.py

-            rect = qImage.rect()
-            p = QtGui.QPainter(self)
-            # reset the image area of the canvas to be the back-ground color
-            p.eraseRect(rect)


This is important if users have alpha in the final rgba. If this is not here text appears to 'get thicker' over time as the alpha stacks up.

Good catch, restored the erase in the case where bbox_queue is empty.

tacaswell · 2017-06-19T03:59:57Z

lib/matplotlib/backends/backend_qt5agg.py

@@ -244,14 +190,11 @@ def __init__(self, figure):
        if DEBUG:
            print('FigureCanvasQtAgg: ', figure)
        super(FigureCanvasQTAgg, self).__init__(figure=figure)
-        self._drawRect = None
-        self.blitbox = []


This going private needs to be documented and gracefully deprecated. Users who use blitting are extra clever and are likely abusing this.

done (modulo comment below)

tacaswell · 2017-06-19T04:01:38Z

lib/matplotlib/backends/backend_qt5agg.py

+            # Adjust the buf reference count to work around a memory
+            # leak bug in QImage under PySide on Python 3.
+            if QT_API == 'PySide' and six.PY3:
+                ctypes.c_long.from_address(id(buf)).value = 1


I am curious what the history on why we had this done two different ways was.

tacaswell · 2017-06-19T04:10:30Z

lib/matplotlib/backends/backend_cairo.py

@@ -56,12 +59,6 @@
 from matplotlib.transforms   import Bbox, Affine2D
 from matplotlib.font_manager import ttfFontProperty

-# Image::color_conv(format) for draw_image()


This came in via 9cb48c1 and it's use was removed in #5718 and it's backports.

tacaswell · 2017-06-19T04:12:34Z

lib/matplotlib/backends/backend_cairo.py

@@ -509,7 +476,7 @@ def print_svg(self, fobj, *args, **kwargs):
    def print_svgz(self, fobj, *args, **kwargs):
        return self._save(fobj, 'svgz', *args, **kwargs)

-    def _save (self, fo, format, **kwargs):
+    def _save(self, fo, fmt, **kwargs):


I will grumble a bit about this, but it is OK. It is a private API and the change is to avoid shadowing a builtin.

tacaswell · 2017-06-19T04:15:46Z

I think the biggest concern I have with this is the dropping of eraserect and the little endian support, the rest of it seems non-controversial on a first read.

anntzer · 2017-06-19T04:52:45Z

You mean blitbox? I made it a deprecated property, do you want an API change there? (tbh I am not sure how that would help anyone, there is simply no replacement for it (other than calling .blit perhaps)) eraseRect was already private.

anntzer · 2017-06-19T05:11:18Z

New version pushed.

tacaswell · 2017-06-19T23:51:56Z

lib/matplotlib/backends/backend_qt5agg.py

@@ -75,9 +63,14 @@ def drawRectangle(self, rect):
            self._drawRect = None
        self.update()

+    @property
+    @cbook.deprecated("2.0")


Should this be 2.1?

efiring · 2017-06-20T09:20:19Z

On 2017/06/20 12:51 AM, Thomas A Caswell wrote: ***@***.**** commented on this pull request. ------------------------------------------------------------------------ In lib/matplotlib/backends/backend_qt5agg.py <#8772 (comment)>: > @@ -86,88 +84,37 @@ def paintEvent(self, e): if not hasattr(self, 'renderer'): return - if DEBUG: - print('FigureCanvasQtAgg.paintEvent: ', self, - self.get_width_height()) - - if len(self.blitbox) == 0: - # matplotlib is in rgba byte order. QImage wants to put the bytes - # into argb format and is in a 4 byte unsigned int. Little endian - # system is LSB first and expects the bytes in reverse order - # (bgra). - if QtCore.QSysInfo.ByteOrder == QtCore.QSysInfo.LittleEndian: fair. There are not many little-endian systems left in practice (and they tend to be big main frame type systems these days).

For the record, I think you meant "big-endian".

tacaswell · 2017-07-01T19:17:06Z

power-cycled to retrigger CI.

tacaswell · 2017-07-01T20:00:09Z

lib/matplotlib/backends/backend_qt5agg.py

+
+        if self._bbox_queue:
+            bbox_queue = self._bbox_queue
+            self._bbox_queue = []


Do not need this line, unconditionally reset below.

tacaswell · 2017-07-01T20:04:01Z

lib/matplotlib/backends/backend_qt5agg.py

@@ -196,7 +141,7 @@ def __draw_idle_agg(self, *args):
            self._agg_draw_pending = False
            return
        try:
-            FigureCanvasAgg.draw(self)
+            super(FigureCanvasQTAggBase, self).draw()


This can be simplified to self.draw()

tacaswell · 2017-07-01T20:06:32Z

lib/matplotlib/backends/backend_cairo.py

-        self.ctx.set_line_width (self.renderer.points_to_pixels(w))
-
-
-def new_figure_manager(num, *args, **kwargs): # called by backends/__init__.py


I think this change leaked when you split the changes up?

tacaswell

Left 3 small comments, ok with this going in with out them being addressed.

anntzer · 2017-07-01T22:50:46Z

Comments handled.

Follow a similar architecture to the Gtk3Agg backend: a default draw can be implemented as a blitting of the entire canvas. Also unify the Qt4Agg and Qt5Agg canvas classes a bit more.

The removed comment in backend_agg regarding rgb came in in 281b7b7, but is now irrelevant.

anntzer · 2017-07-09T07:10:08Z

Last commit removes some old comments, rewords some docstrings, reformats a few lines.

This was referenced Jun 18, 2017

Backend class for better code reuse between backend modules #8773

Merged

Cairo rendering for Qt5 #8771

Closed

tacaswell reviewed Jun 19, 2017

View reviewed changes

tacaswell added this to the 2.1 (next point release) milestone Jun 19, 2017

anntzer force-pushed the backends-cleanup branch from 0194395 to 1f50a0f Compare June 19, 2017 05:11

anntzer force-pushed the backends-cleanup branch from 1f50a0f to 0fd5b04 Compare June 19, 2017 05:48

tacaswell reviewed Jun 19, 2017

View reviewed changes

anntzer force-pushed the backends-cleanup branch from 0fd5b04 to 6ce2f06 Compare June 20, 2017 00:04

anntzer force-pushed the backends-cleanup branch from 6ce2f06 to 22457f3 Compare June 27, 2017 09:13

tacaswell closed this Jul 1, 2017

tacaswell reopened this Jul 1, 2017

tacaswell reviewed Jul 1, 2017

View reviewed changes

tacaswell approved these changes Jul 1, 2017

View reviewed changes

anntzer force-pushed the backends-cleanup branch from 22457f3 to f879b2b Compare July 1, 2017 22:50

Unify QtAgg nonblitting and blitting code paths.

65a0ccc

Follow a similar architecture to the Gtk3Agg backend: a default draw can be implemented as a blitting of the entire canvas. Also unify the Qt4Agg and Qt5Agg canvas classes a bit more.

anntzer added 3 commits July 8, 2017 20:31

Remove unused _needs_redraw from gtk3 backends.

05e5ae1

Minor cleanups.

6d2380e

And some more cleanups.

79cfe14

The removed comment in backend_agg regarding rgb came in in 281b7b7, but is now irrelevant.

anntzer force-pushed the backends-cleanup branch from f879b2b to 79cfe14 Compare July 9, 2017 07:06

tacaswell merged commit d412228 into matplotlib:master Jul 9, 2017

anntzer deleted the backends-cleanup branch July 9, 2017 19:44

		self.ctx.set_line_width (self.renderer.points_to_pixels(w))


		def new_figure_manager(num, args, *kwargs): # called by backends/__init__.py

Uh oh!

Backends cleanup #8772

Backends cleanup #8772

Uh oh!

Conversation

anntzer commented Jun 18, 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!

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!

tacaswell Jun 19, 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!

tacaswell commented Jun 19, 2017

Uh oh!

anntzer commented Jun 19, 2017

Uh oh!

anntzer commented Jun 19, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

efiring commented Jun 20, 2017 via email

Uh oh!

tacaswell commented Jul 1, 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!

tacaswell left a comment

Choose a reason for hiding this comment

Uh oh!

anntzer commented Jul 1, 2017

Uh oh!

anntzer commented Jul 9, 2017

Uh oh!

Uh oh!

tacaswell Jun 19, 2017 •

edited

Loading