As the replay of the recording surface applies the device_transform of
the matrix once again to all its operations, we end up with a repeated
transform through the source matrix of the recording surface. We need to
remove one of these, and the easiest way to do that appears to be to
undo the application to the source matrix.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When creating a transformed pattern we must apply the device
transform *before* the transform set on the pattern itself, otherwise
e.g. its translation will not be affected by the device scale.
We also fix up the device_transform related handling in
_cairo_default_context_pop_group(). With a device scale we can
no longer just use the device_transform_inverse to unset the
device offset for the extents, so we make that a simple translate
instead.
We also remove some weird code that tries to handle the device
transform but seems unnecessary (maybe a workaround for applying
the device transform in the wrong order?). With that code removed
things work fine, but with it things get translated wrongly when
there is a scale.
If we do this in surface it will be applied twice then
we chain to a different surface, like e.g. a subsurface.
We also remove a hack in cairo-surface-wrapper where it compensated
for the device scale not being applied.
v2: Compute the backend CTM in ensure_scaled_font().
I had a look at how complex would it be to add correct downscaling to
Cairo now that Pixman supports convolution filters. It turns out it
this is rather easy. Here is an initial, minimal attempt. It uses
convolution filters only if the image is being downscaled by more than
half a pixel in at least one dimension.
Some discussion:
1. The sampling and reconstruction kernels are picked in a way that
gives comparable quality when upscaling and downscaling. I paired box
sampling with bilinear reconstruction and impulse (point) sampling
with box reconstruction. This gives the expected result for NEAREST
filter. BEST filter uses Lanczos3 for both kernels.
> Do we need to use a reconstruction filter for NEAREST at all? Or maybe
> differentiate between NEAREST and FAST in that case?
If impulse (point) sampling is used, there must be some reconstruction
filter, otherwise no image is produced. That's because the sampling
grid does not match the data grid, and since there is no
reconstruction filter, values between data points are undefined. The
alternative is to use box sampling + no reconstruction.
2. Subsampling bits are always set to 1, since this doesn't seem to
affect quality at all.
3. I am not sure whether this code works correctly for matrices with a
skew component. It should be OK for any combination of scale, rotation
and translation.
4. This patch causes new failures in the test suite:
- recording-surface*: possibly an effect of improved quality.
- surface-pattern-scale-down*, surface-pattern-big-scale-down: the
reference images should be updated.
- pthread-same-source: I have no idea why this is failing, since this
test shouldn't even trigger the new code.
- large-source-roi: this test attempts to downscale an image which is
30000 pixels wide down to 7 pixels. The filter parameters seem to be
created correctly, but they might trigger an overflow somewhere in the
convolution code; the output rectangle is white instead of red, as if
nothing was drawn.
- device-offset-scale: there are subtle differences which look like
convolution-related smoothing; I'm not sure whether this is OK or not.
Based on a patch and analysis by Michael Henning.
When we create a similar-image surface for win32, we set up a couple of
back references from the image to the win32 surface, and vice versa. We
need to be careful when decoupling the reference cycle to avoid chasing
around the loop upon destruction. Currently we handled destroying the
similar-image via the parent win32 surface, but similar precaution is
required when destroying the surface via the similar-image.
Reported-by: Michael Henning <drawoc@darkrefraction.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=63787
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When creating a texture surface that is larger than the maximum
framebuffer or texture dimensions of the context, return a surface in
error. Previously the code failed an assertion, but this prevents an
application from easily detecting when to fall back.
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
This fixes a crash in the api-special-cases with xlib-xcb when calling
cairo_clip_extents() on a context that refers to a finished surface.
The crash was a simple NULL pointer dereference, because the underlying xcb
surface that was used in xlib-xcb was gone and set to NULL already.
Signed-off-by: Uli Schlachter <psychon@znc.in>
Make cairo_push_group() fail when the context's target surface is finished.
This fixes the api-special-cases for the xcb backend:
Detected error during xcb run: error=9, seqno=0x13c, major=53, minor=0
The problem was that the Pixmap for the cairo surface was already freed and
cairo still tried to use it again as the drawable in a CreatePixmap request.
Signed-off-by: Uli Schlachter <psychon@znc.in>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Finished surfaces and surfaces with an error status must not be usable anymore,
so refuse to work on them.
This improves the result for api-special-cases.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=68014
Signed-off-by: Uli Schlachter <psychon@znc.in>
This adds code to the api-special-cases test which also tests the behavior of
cairo when the cairo context or the surface that is target is in an error state
or finished. These new tests call into all public entry points defined in
cairo.h which receive a cairo_t * as their first argument.
Currently this causes a new crash in the testsuite:
cairo-surface.c:394:
_cairo_surface_begin_modification: Assertion `! surface->finished' failed.
Reported-by: christophe.troestler@umons.ac.be
References: https://bugs.freedesktop.org/show_bug.cgi?id=68014
Signed-off-by: Uli Schlachter <psychon@znc.in>
The macro benchmarks were moved to a separate repository some time ago,
but the perf README still refers to these tests as if they were still
present, which may lead to some confusion. Instead, consolodate the
macro benchmark documentation with the macro benchmarks, and focus this
README on just the (still in tree) micro-benchmarks.
Signed-off-by: Bryce Harrington <b.harrington@samsung.com>
Signed-off-by: Uli Schlachter <psychon@znc.in>
When the latin subset contains glyphs that use the seac operator to
combine two glyphs, additional non winansi glyphs are added to the
subset. These extra glyphs do not have a winansi name so they can't
be renamed.
Based on a patch by Salvador Ortiz.
Bug 67324
When blitting the framebuffer during transitions to and from
multi-sampling mode, we need to disable the stencil and scissor test so
that the entire surface is preserved. This fixes the bitmap-font test
for the MSAA compositor.
If subpixel rendering is enabled, but FT returns a 8bit gray bitmap
(perhaps because the font has 8bit embedded bitmaps) we were hitting
the assertions because the assumptions made were wrong. Fix up.
In some obscure conditions that I don't really understand, the image compositor
did not finish a span renderer that it created. This could then cause the last
row of the span to be ignored.
Fixes: clip-complex-bug61492
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=61592
Signed-off-by: Uli Schlachter <psychon@znc.in>
Commit d7f5a1bec fixed a bug. This caused 12 new test failures for the
test-traps test target:
caps-tails-curve degenerate-arc degenerate-path joins subsurface
subsurface-scale twin twin-antialias-gray twin-antialias-mixed
twin-antialias-none twin-antialias-subpixel user-font
Most of these are indeed (new?) bugs. However, caps-tails-curve actually started
producing the expected result and the reference image just wrongly captures the
old state of things.
At the time of that commit, just taking the output from test-traps as the new
reference image works fine for all backends. However, with current git,
something introduced more antialiasing noise and now test-traps changed again
while cairo-xcb stayed with the old result. Thus, we also need a new reference
image to fix this test.
(The wrong reference images come from commit 8488ae02 which turned test-traps'
results into reference images)
Signed-off-by: Uli Schlachter <psychon@znc.in>
Where a content specific reference image exists, prefer to have both
content reference images (i.e. both argb32.ref and rgb24.ref) rather
than a mix of .ref and argb32/rgb24.
Courtesy of the improved check-ref-dups written by Bryce Harrington:
Running make check on the codebase (with default configuration) with the
redundant images removed produces essentially the same test results:
Before
------
Tests run: 13687
Passed: 9216
Failed: 3566
Expected Failed: 312
Error: 1
Crashed: 17
Untested: 575
Total: 13687
After
-----
Tests run: 13689
Passed: 9216
Failed: 3566
Expected Failed: 312
Error: 1
Crashed: 19
Untested: 575
Total: 13689
(with the exception being the pthread tests misbehaving between runs)
These tests use reference images somewhat differently from other tests,
so treat them as special cases and avoid recommending deleting any of
their files.
Add TODO's to each test to rework them to be more consistent with other
tests.
Signed-off-by: Bryce Harrington <b.harrington@samsung.com>
perceptualdiff.c:55:19: warning: comparison between signed and unsigned
integer expressions [-Wsign-compare]
perceptualdiff.c:60:16: warning: comparison between signed and unsigned
integer expressions [-Wsign-compare]
Signed-off-by: Bryce Harrington <b.harrington@samsung.com>
The current sha1sum-based file checker does a blanket comparison of all
files with each other, which is fast but not directly helpful since it
doesn't distinguish between sibling files (which are allowable to be
duplicates). Also, it ignores files that may have byte differences
(such as PNG header differences) but are otherwise pixel-identical.
This patch replaces the sha1sum-based checker with one that accounts for
the default fallback path and uses perceptualdiff to compare files that
are bytewise different and verify whether they actually are different.
The intention is that the output of this command can be directly used to
remove redundant files, e.g.:
cd test
make check-ref-dups | cut -d' ' -f2 | \
while read f; do git rm "reference/$f"; done
This should have no impact on make check's results (modulo any tests
that behave erratically), and will help trim down the size of the
tarball.
Signed-off-by: Bryce Harrington <b.harrington@samsung.com>
[ickle: rebase and add the suggested command to the Makefile]
cmp runs faster than perceptualdiff, and catches files that are exact
copies of the reference image. We still use perceptualdiff for catching
files that aren't bytewise identical, but are still identical at the
pixel level.
Signed-off-by: Bryce Harrington <b.harrington@samsung.com>
This script requires the perceptualdiff program, which can be built as
follows:
cd test/pdiff && make perceptualdiff
The script's output provides a list of target-specific or
format-specific images that are identical to their more generic
reference files, and thus are redundant and can be safely removed from
the archive without altering any test behaviors.
Signed-off-by: Bryce Harrington <b.harrington@samsung.com>
[ickle: applied Behdad's suggestion of renaming the script check-refs.sh]
The check-ref-dups target in Makefile.am was not updated with the new
path when the reference images were moved from test/ to
test/reference/. Now it produces output properly again.
Signed-off-by: Bryce Harrington <b.harrington@samsung.com>
The lookup of the string names has significant overhead, which is why GL
gives you glGetUniformLocation so that you reference uniforms by
constant integers in your high performance path.
Reduces cairo-perf-trace runtime of firefox-planet-gnome by 1.06767% +/-
0.289265% (n=72) on my IVB macbook air.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
