Faster path drawing for the cairo backend (cairocffi only) by anntzer · Pull Request #8787 · matplotlib/matplotlib · GitHub

When using cairocffi only:

Accelerate Path drawing by directly constructing a cairo_path_t using numpy operations rather than repeatedly calling the cairo API to add one point at a time.
Accelerate PathCollection drawing by concatenating Paths that have the same drawing parameters (color, etc., i.e. GraphicsContext attributes).
Minor additional performance improvements (remove an unneeded ndarray copy and an unneeded context save/restore pair).

Overall, this brings the performance of examples/mplot3d/wire3d_animation.py (used as a microbenchmark) from ~8.3fps to ~11.6fps on the same computer (just optimizing Paths instead of PathCollections gives ~10.5fps), using the gtk3cairo backend. As a comparison point, the gtk3agg backend renders ~16.2fps for that example.

From profiling, I believe that further improvements (on that same microbenchmark) could be achieved by improving the text drawing (currently, we use the "toy" API, which is unfortunately the only one exposed by cairocffi so far; additionally, I am not certain whether ft2font exposes pointers to the underlying freetype structures).

Or (for other performance tests) draw_markers could probably also benefit from a similar approach as this PR.

Not sure why this is different than before?

previous version was actually incorrect (see basically https://en.wikipedia.org/wiki/B%C3%A9zier_curve#Degree_elevation)..

Maybe if at top-level would be quicker/simpler?

By top-level, I meant if HAS_CAIRO_CFFI: _convert_paths = _convert_paths_fast ...

s/is for/for/

Could be += to make one line?

transforms are mutable but don't implement iadd (so += actually doesn't modify the transform in place), but I was too lazy to check and want to make clear that we're indeed not mutating the transform.

Also +=?

as above

This fits on one line, no?

No underscore is needed?

Actually no, I'm just copying the __dict__ of the original gc, and (kind of an implementation detail...) GraphicsContextBase stores "property" foo (accessed by get/set_foo) in ._foo). Not very elegant I admit.

Why do you need the second Affine2D?

I don't.

Do we need to worry about draw order and such here?

The draw order is maintained, it's just that I batch the draws that use the same gc together.

Could be re-arranged slightly to keep the int and float bits together.

Not sure what you meant?

Well, you start with a buffer, make an int version, then a float version, then a mask (which is for float). But then go back to filling the int version, modifying the mask (which is float stuff again), then fill the float version. So I mean more like:

    buf = np.empty(cairo_num_data * 16, np.uint8)
    as_int = np.frombuffer(buf.data, np.int32)
    as_int[::4][cairo_type_positions] = codes
    as_int[1::4][cairo_type_positions] = cairo_type_sizes
    as_float = np.frombuffer(buf.data, np.float64)
    mask = np.ones_like(as_float, bool)
    mask[::2][cairo_type_positions] = mask[1::2][cairo_type_positions] = False
    as_float[mask] = vertices.ravel()

Comments addressed.

comments addressed

Generally, I don't see a problem with this change, but using your microbenchmark, I can't seem to reproduce the speedup you see:

$ python3 examples/mplot3d/wire3d_animation.py
Matplotlib version: 2.1.1.post1307.dev0+g84008ea05
Backend: TkAgg
Average FPS: 19.965329

$ python3 examples/mplot3d/wire3d_animation.py
/usr/lib/python3.6/site-packages/cairocffi/surfaces.py:651: UserWarning: implicit cast from 'char *' to a different pointer type: will be forbidden in the future (check that the types are as you expect; use an explicit ffi.cast() if they are correct)
  ffi.cast('char*', address), format, width, height, stride)
Matplotlib version: 2.1.1.post1307.dev0+g84008ea05
Backend: GTK3Agg
Average FPS: 20.217351

$ python3 examples/mplot3d/wire3d_animation.py
Matplotlib version: 2.1.1.post1307.dev0+g84008ea05
Backend: agg
Average FPS: 22.100502

$ python3 examples/mplot3d/wire3d_animation.py
Matplotlib version: 2.1.1.post1307.dev0+g84008ea05
Backend: GTK3Cairo
Average FPS: 12.131168

$ python3 examples/mplot3d/wire3d_animation.py
Matplotlib version: 2.1.1.post1307.dev0+g84008ea05
Backend: cairo
Average FPS: 70.399400

vs.

$ python3 examples/mplot3d/wire3d_animation.py
Matplotlib version: 2.1.1.post18.dev0+gebcfd908c
Backend: TkAgg
Average FPS: 18.767531

$ python3 examples/mplot3d/wire3d_animation.py
/usr/lib/python3.6/site-packages/cairocffi/surfaces.py:651: UserWarning: implicit cast from 'char *' to a different pointer type: will be forbidden in the future (check that the types are as you expect; use an explicit ffi.cast() if they are correct)
  ffi.cast('char*', address), format, width, height, stride)
Matplotlib version: 2.1.1.post18.dev0+gebcfd908c
Backend: GTK3Agg
Average FPS: 19.497821

$ python3 examples/mplot3d/wire3d_animation.py
Matplotlib version: 2.1.1.post18.dev0+gebcfd908c
Backend: agg
Average FPS: 21.796568

$ python3 examples/mplot3d/wire3d_animation.py
Matplotlib version: 2.1.1.post18.dev0+gebcfd908c
Backend: GTK3Cairo
Average FPS: 12.018317

$ python3 examples/mplot3d/wire3d_animation.py
Matplotlib version: 2.1.1.post18.dev0+gebcfd908c
Backend: cairo
Average FPS: 68.990914

Are you certain your gtk3cairo is using cairocffi rather than pycairo (which is not accelerated by this PR)?
I just tried and can still get the difference, ~8.7fps vs 12.2fps.

OK, I'm not really sure what happened, but I just tried again and I get 12.705427 vs. 17.470702 now.

Not a cario expert, but look good to me.

Not really clear with the naming of convert_paths. To me convert feels like it should return the converted object. Maybe apply_paths? If that's not good, please add a short docstring to the convert functions.

flatten() always makes a copy, whereas ravel() does not.

The removed pair of ctx.save and ctx.restore was clearly unnecessary (the outer one is still there, and the font is reset at each loop iteration).

Improves the performance of mplot3d/wire3d_animation on the gtk3cairo backend from ~8.3fps to ~10.5fps (as a comparison, gtk3agg is at ~16.2fps).

Further increase the performance of mplot3d/wire3d_animation on the gtk3cairo backend from ~10.5fps to ~11.6fps (as a comparison, gtk3agg is at ~16.2fps).

For the slow code path, implement the degree elevation formula. For the fast code path, the path cleaner was already handling this for us, converting everything to lines.

Renamed to _append_path, which matches the name of the function in the cairo API: https://cairographics.org/manual/cairo-Paths.html#cairo-append-path

Would it help you if pycairo had something like cairo.Path.create_for_data(buffer, num_data) ?

I moved my efforts to just doing everything in C(++) (https://github.com/anntzer/mplcairo) but I guess such a method can't hurt in general...
Perhaps also worth (for you) to discuss this with cairocffi to present the same API in both projects.

I moved my efforts to just doing everything in C(++) (https://github.com/anntzer/mplcairo) but I guess such a method can't hurt in general...

ok

Perhaps also worth (for you) to discuss this with cairocffi to present the same API in both projects.

yeah, but it's unmaintained atm

anntzer added the backend: cairo label Jun 22, 2017

anntzer force-pushed the cairo-fast-path branch 3 times, most recently from fe66d3b to 37775fb Compare June 22, 2017 22:11

tacaswell added this to the 2.2 (next next feature release) milestone Jun 24, 2017

anntzer force-pushed the cairo-fast-path branch from 706de3a to abce702 Compare June 27, 2017 09:16

anntzer mentioned this pull request Aug 28, 2017

PDFPages Errors on savefig (missing encode) #9112

Closed

anntzer force-pushed the cairo-fast-path branch from abce702 to 241a306 Compare November 4, 2017 07:06

anntzer mentioned this pull request Jan 16, 2018

qt{4,5}cairo backend: the minimal version. #10210

Merged

6 tasks

QuLogic reviewed Feb 3, 2018

View reviewed changes

anntzer force-pushed the cairo-fast-path branch from 241a306 to f1824ee Compare February 3, 2018 13:16

anntzer force-pushed the cairo-fast-path branch 2 times, most recently from a103612 to 245f5de Compare February 6, 2018 08:24

anntzer force-pushed the cairo-fast-path branch from 245f5de to 4a9f752 Compare February 6, 2018 08:27

QuLogic approved these changes Feb 15, 2018

View reviewed changes

anntzer modified the milestones: needs sorting, v3.0 Feb 26, 2018

timhoffm approved these changes May 6, 2018

View reviewed changes

anntzer added 7 commits May 6, 2018 14:09

cairo: Don't copy the ravelled image.

3edc982

flatten() always makes a copy, whereas ravel() does not.

cairo: save a pair of ctx.save/ctx.restore.

b549d12

The removed pair of ctx.save and ctx.restore was clearly unnecessary (the outer one is still there, and the font is reset at each loop iteration).

Faster path drawing with cairocffi.

8a07eef

Improves the performance of mplot3d/wire3d_animation on the gtk3cairo backend from ~8.3fps to ~10.5fps (as a comparison, gtk3agg is at ~16.2fps).

Implement draw_path_collection for cairo.

52ed746

Further increase the performance of mplot3d/wire3d_animation on the gtk3cairo backend from ~10.5fps to ~11.6fps (as a comparison, gtk3agg is at ~16.2fps).

Document a bit the cairo fast path.

5312461

Actually support quadratic Beziers.

1a2848d

For the slow code path, implement the degree elevation formula. For the fast code path, the path cleaner was already handling this for us, converting everything to lines.

Address review comments.

aca804f

anntzer added 2 commits May 6, 2018 14:09

Avoid triggering a DeprecationWarning.

7a727a1

Rebase for mpl3.

abbcb3e

anntzer force-pushed the cairo-fast-path branch from 4a9f752 to abbcb3e Compare May 6, 2018 21:20

timhoffm merged commit 0d71eba into matplotlib:master May 6, 2018

anntzer deleted the cairo-fast-path branch May 6, 2018 23:20

anntzer mentioned this pull request Nov 15, 2018

Cairo backend: Fix alpha render of collections #12774

Merged

5 of 6 tasks

+                              cur = ctx.get_current_point()
+                              ctx.curve_to(
+                                  *np.concatenate([cur / 3 + points[:2] * 2 / 3,
+                                                   points[:2] * 2 / 3 + points[-2:] / 3]))

-                      self.convert_path(
-                          ctx, marker_path, marker_trans + Affine2D().scale(1.0, -1.0))
+                      _convert_path(
+                          ctx, marker_path, marker_trans + Affine2D().scale(1, -1))

+                          vars(gc).update(gc_vars)
+                          for k, v in gc_vars.items():
+                              try:
+                                  getattr(gc, "set" + k)(v)

+                              antialiaseds, urls, offset_position):
+                          path, transform = path_id
+                          transform = (Affine2D(transform.get_matrix()).translate(xo, yo)
+                                       + Affine2D().scale(1, -1).translate(0, self.height))

+                          if new_key == reuse_key:
+                              grouped_draw.append((path, transform))
+                          else:
+                              _draw_paths()



		def _convert_paths(ctx, paths, transforms, clip=None):
		return (_convert_paths_fast if HAS_CAIRO_CFFI else _convert_paths_slow)(

+              def _convert_paths_fast(ctx, paths, transforms, clip=None):
+                  # We directly convert to the internal representation used by cairo, for
+                  # which ABI compatibility is guaranteed.  The layout is for each item is

                       transform = (transform
-                                   + Affine2D().scale(1.0, -1.0).translate(0, self.height))
+                                   + Affine2D().scale(1, -1).translate(0, self.height))

Faster path drawing for the cairo backend (cairocffi only) by anntzer · Pull Request #8787 · matplotlib/matplotlib · GitHub

Repository navigation

Faster path drawing for the cairo backend (cairocffi only) - #8787

Faster path drawing for the cairo backend (cairocffi only)#8787timhoffm merged 9 commits intomatplotlib:mastermatplotlib/matplotlib:masterfrom anntzer:cairo-fast-pathanntzer/matplotlib:cairo-fast-pathCopy head branch name to clipboard

Conversation

anntzer commented Jun 22, 2017 • edited 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!

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!

anntzer Feb 3, 2018 • 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!

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 Feb 3, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

anntzer commented Feb 6, 2018

Uh oh!

QuLogic commented Feb 8, 2018

Uh oh!

anntzer commented Feb 8, 2018

Uh oh!

QuLogic commented Feb 15, 2018

Uh oh!

timhoffm left a comment

Choose a reason for hiding this comment

Uh oh!

anntzer commented May 6, 2018

Uh oh!

lazka commented May 28, 2018

Uh oh!

anntzer commented May 28, 2018

Uh oh!

lazka commented May 28, 2018

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Faster path drawing for the cairo backend (cairocffi only)#8787
timhoffm merged 9 commits into
matplotlib:masterfrom
anntzer:cairo-fast-path

anntzer commented Jun 22, 2017 •
edited
Loading

anntzer Feb 3, 2018 •
edited
Loading

anntzer commented Feb 3, 2018 •
edited
Loading