Can you give a bit of a guide for how to test this worked? I think the general idea is great - to make the mac backend as similar as possible to the others....

I was testing on the examples from the linked issues as they had pretty good minimal code samples. All of them basically didn't show any updates (redraw the entire canvas background which was white, rather than blitting and updating only the new portion of the figure).

#10980 (Should rotate around)

frommpl_toolkits.mplot3dimportaxes3dimportmatplotlib.pyplotaspltfig=plt.figure()ax=fig.add_subplot(111,projection='3d')# load some test data for demonstration and plot a wireframeX,Y,Z=axes3d.get_test_data(0.1)ax.plot_wireframe(X,Y,Z,rstride=5,cstride=5)# rotate the axes and updateforangleinrange(0,360):ax.view_init(30,angle)plt.draw()plt.pause(.001)

#17445 (Should not print "Drawing!" 40+ times and also update the arrow animation)

importmatplotlib.pyplotaspltfrommatplotlibimportanimationimportnumpyasnpclassBaseVisualizer():def__init__(self,**kwargs):self.__iteration_number=kwargs.get('iteration_number',1)self.__intervals=1self.__interval_ms=kwargs.get('interval',2000)self._marker_size=4self._index=0self._positions=np.zeros(2)self._velocities=np.ones(2)self.__frame_interval=50# msdefreplay(self):self._fig=plt.figure()self._fig.canvas.mpl_connect('draw_event',lambdaev:print('Drawing!'))self._fig.stale_callback=Noneax=self._fig.add_subplot(1,1,1,label='BaseAxis')# Plot all particle posself.__particles=ax.scatter([], [],marker='o',zorder=2,color='red')# Plot all velocitiesself.__particle_vel=ax.quiver([], [], [], [],angles='xy',scale_units='xy',scale=1)self.__rectangle=plt.Rectangle([0,0],1,1,ec='none',lw=2,fc='none')ax.add_patch(self.__rectangle)# Amount of frames to play considering one interval should last __interval ms.frames=int(self.__intervals*self.__interval_ms/self.__frame_interval)# iteration_number+1 for initialization frame_=animation.FuncAnimation(self._fig,self._animate,frames=frames,interval=self.__frame_interval,blit=True,repeat=False,fargs=[frames])plt.show()def_animate(self,i:int,frames:int):"""        Animation function for animations. Only used for FuncAnimation        """self._marker_size=int(50*self._fig.get_figwidth()/self._fig.dpi)# Calculate the scale to apply to the data in order to generate a more dynamic visualizationscale=i/ (frames/self.__intervals)# Calculate scaled position and velocitypos_scaled=self._positions+scale*self._velocitiesvel_scaled= (1-scale)*self._velocities# Update the particle positionself.__particles.set_offsets(pos_scaled)self.__particles.set_sizes([self._marker_size**2])# Update the velocitiesax=self._fig.gca(label='BaseAxis')self.__particle_vel=ax.quiver(pos_scaled[0],pos_scaled[1],vel_scaled[0],vel_scaled[1],angles='xy',scale_units='xy',scale=1,color='#CFCFCF',width=self._marker_size*0.001)returnself.__particles,self.__particle_velif__name__=='__main__':vis=BaseVisualizer()vis.replay()

#17642 (A sine wave should move across the screen)

importmatplotlib.pyplotaspltimportnumpyasnpx=np.linspace(0,2*np.pi,100)fig,ax=plt.subplots()# animated=True tells matplotlib to only draw the artist when we# explicitly request it(ln,)=ax.plot(x,np.sin(x),animated=True)# make sure the window is raised, but the script keeps goingplt.show(block=False)# stop to admire our empty window axes and ensure it is rendered at# least once.## We need to fully draw the figure at its final size on the screen# before we continue on so that :#  a) we have the correctly sized and drawn background to grab#  b) we have a cached renderer so that ``ax.draw_artist`` works# so we spin the event loop to let the backend process any pending operationsplt.pause(0.1)# get copy of entire figure (everything inside fig.bbox) sans animated artistbg=fig.canvas.copy_from_bbox(fig.bbox)# draw the animated artist, this uses a cached rendererax.draw_artist(ln)# show the result to the screen, this pushes the updated RGBA buffer from the# renderer to the GUI framework so you can see itfig.canvas.blit(fig.bbox)forjinrange(100):# reset the background back in the canvas state, screen unchangedfig.canvas.restore_region(bg)# update the artist, neither the canvas state nor the screen have changedln.set_ydata(np.sin(x+ (j/100)*np.pi))# re-render the artist, updating the canvas state, but not the screenax.draw_artist(ln)# copy the image to the GUI state, but screen might not changed yetfig.canvas.blit(fig.bbox)# flush any pending GUI events, re-painting the screen if neededfig.canvas.flush_events()# you can put a pause in if you want to slow things down# plt.pause(.1)

I can confirm this fixes the attached examples. I don't think I have the expertise to thoroughly review the code changes, but within what I know they look fine. Is there a way to add a blitting test for the macosx backend to the test suite?

Is there a way to add a blitting test for the macosx backend to the test suite?

This is definitely a good idea, but as far as I can tell if I try to save anything during the process, the buffer gets updated and the draw does actually happen, so the old versions would actually still appear to work from a test perspective, even though they don't on the screen. I tried grabbing the buffer during one of the animations, and the buffer is getting updated I believe, it is just Mac isn't drawing the buffer to the screen. If anyone has a suggestion for how to do the final screen buffer grab I'm happy to implement that.

Here is the quick check I was trying comparing new/old buffer bytes before/after blitting or animating:

new_bytes=bytes(fig.canvas.buffer_rgba())print(new_bytes==orig_bytes)# False on both main and this PR

@dstansby, did the final example with the sine wave work for you? When I go back to that today I don't get the flush_events to work as expected, so I may have had something crossed up in my environments. I think I may also need to bring over the commit from#21768 after all.

OK, I must have had something mixed up in my test cases before and thisdoes need the update to flush_events so that the macosx event loop has time to process the updates. I brought the commit from#21790 over here, but I'm happy to reopen that one again if it would be easier consider them separately.

I just pushed up a new commit that I think addresses thedraw_idle callback. I tried to add a singleshot timer and start that without needing to stop the timer, but unfortunately that segfaults on the example from#17445. There is something about threading going on in there (if I removePy_BEGIN_ALLOW_THREADS it works as expected), but I couldn't figure out how to resolve it at the C level. It also only affects the single-shot timers, if I make it a repeating timer then everything works just fine again. So, my guess is there is some race condition going on with the deallocation between threads and when the timers get released.

I was able to get around this by stopping the timer at the Python level after firing. I'm not sure that is the most elegant way around this, so I'm open to suggestions if anyone has other ideas for how to approach the runloop singleshot timers here.

I have now added a test to verify that we aren't getting multiple draw_events emitted from the backends. I added it intotest_backends_interactive as that is where the other show() based tests are, but I could see it going over totest_animation too. "gtk3agg" was getting multiple draw events for me locally, and "gtk3cairo" was complaining about blitting on the canvas, so I decided to skip gtk3 for now with a FIXME to look into that later.

It looks like the test I added is not happy on some architectures, before I go banging on CI I figured I would see what others would prefer. I see a few options:

1. Make it a macosx only test since that is what this PR is focused on and fixing
2. Increase the timeouts/pauses even more to allow the first draw events to flush out (I think that is why tkagg is failing, as it passes for me locally), this comes at the cost of increased test time.
3. Start removing cases from the tests here (wx[agg], and tkagg), comes at the cost of less test coverage. wx/wxagg I'm not exactly sure what is going on there, if a longer pause would help or if the blitting is broken there too.
4. Find a better way to test this (would need some suggestions here!)

A possible idea:frames can be a generator, make it yield a few times and then schedule a single-shot timer to close the figure. You can then doplt.show(block=True), rely on the generator to close the window and exit the event loop and not have to have an arbitrary timeout.

1. I apologise for asking if a test was possible 😄
2. I'm going to unsubscribe, but feel free to ping me when this is ready for review again

Fully avoiding the animation framework and just testing the thing we want to test is amuch better idea than inviting generators to the party!

@dstansby, I think this is good to go for a check from you again. i.e. can you double check me on those examples to make sure I didn't screw something up? Don't apologize for the test request! It is always good to add tests when we can.

@tacaswell, This steals most of your other blitting example for the test, so thanks for that :) Unfortunately, the actual reason for going this direction was because macosx has an issue with single-shot timers and not quitting out of the test on the runners when using the generator/animation approach. Locally it hangs until I move my mouse or do something else interactive on the system, so there is some other issue with macosx event loops and the handoff from Python <-> NSApp which I will try and make a reproducible example for and open an issue.

The examples in#10980 and#17445 work well for me on this branch 👍

Any mac users want to test this out and push it through the finish line?

ping@dstansby who recently reported owning a M1 machine and volunteering to test mac things ;)

Of the examples listed in#21790 (comment), the 1st and 3rd work for me with this branch, but the 2nd does not. It prints "Drawing!" once, and nothing changes in the empty plot that appears.

Thanks for checking@dstansby! That one has a very small marker and thin grey line that are updated, so maybe it just didn't show up as obviously (especially with HiDPI screen)? It does work for me locally copy-pasting that example.

Can you try changing the marker size to make it more obvious:self._marker_size = int(500 * self._fig.get_figwidth() / self._fig.dpi)

If that doesn't work, my guess is this would be an M1 issue then :-/

Thanks for checking! (It was not the easiest example to see) I'm going to self-merge based on the two approvals and this isn't targeting 3.5.x