Bug 84044 - test/cairo-test-suite reports wrong max diff comparing different opaque images
Summary: test/cairo-test-suite reports wrong max diff comparing different opaque images
Status: RESOLVED MOVED
Alias: None
Product: cairo
Classification: Unclassified
Component: general (show other bugs)
Version: unspecified
Hardware: Other All
: medium normal
Assignee: Chris Wilson
QA Contact: cairo-bugs mailing list
URL:
Whiteboard:
Keywords:
Depends on:
Blocks:
 
Reported: 2014-09-18 14:16 UTC by Massimo
Modified: 2018-08-25 13:29 UTC (History)
0 users

See Also:
i915 platform:
i915 features:


Attachments
patch (940 bytes, patch)
2014-09-19 08:35 UTC, Massimo
Details | Splinter Review

Description Massimo 2014-09-18 14:16:04 UTC
In test/buffer-diff.c (line 80), when two pixels differ, all channels
are considered to compute result.max_diff, this means in the case of
rgb24 images also the meaningless alpha channel.

Similarly pdiff_compare in test/pdiff/pdiff.c uses the (non)alpha
channel to unpremultiply the not premultiplied rgb channels, hardly
giving the expected results.

Probably there should be a test-suite for pdiff to check for false
positive/negative.

This results in many image.rgb24 failures that would be accepted
using the same criterion used for image.argb32 target making the
cairo test-suite output quite confusing.

BTW to slightly speed up the test-suite you could write and use a
'pdiff_compare_fast' that returns 1 the first time pixels_failed is
incremented.

The best way to speed up the test-suite is to have all tests PASS though.
Comment 1 Massimo 2014-09-19 08:35:45 UTC
Created attachment 106538 [details] [review]
patch

Running

DISPLAY=:2 time -p make -s test TARGETS=image,xlib,xcb

without/with the attached patch on my laptop the last lines of
test/cairo-test-suite.log change from:

322 Passed, 203 Failed [1 crashed, 8 expected], 29 Skipped
Preamble: 1 failed - create-from-png
image (argb32): 1 crashed! - gl-surface-source
image (argb32): 41 failed - clip-operator extended-blend-alpha exten...
image (rgb24): 1 crashed! - gl-surface-source
image (rgb24): 49 failed - clip-operator extended-blend extended-ble...
xlib (argb32): 1 crashed! - gl-surface-source
xlib (argb32): 104 failed - aliasing arc-looping-dash bug-spline bug...
xlib (rgb24): 1 crashed! - gl-surface-source
xlib (rgb24): 142 failed - aliasing arc-looping-dash bug-spline bug-...
xlib-window (rgb24): 1 crashed! - gl-surface-source
xlib-window (rgb24): 140 failed - aliasing arc-looping-dash bug-spli...
xlib-render-0_0 (rgb24): 1 crashed! - gl-surface-source
xlib-render-0_0 (rgb24): 73 failed - caps-sub-paths clear-source cli...
xlib-fallback (rgb24): 1 crashed! - gl-surface-source
xlib-fallback (rgb24): 113 failed - bug-51910 bug-seams caps caps-jo...
xcb (argb32): 1 crashed! - gl-surface-source
xcb (argb32): 50 failed - bug-spline hatchings huge-radial operator-...
xcb (rgb24): 1 crashed! - gl-surface-source
xcb (rgb24): 99 failed - arc-looping-dash bug-spline bug-51910 bug-s...
xcb-window (rgb24): 1 crashed! - gl-surface-source
xcb-window (rgb24): 97 failed - arc-looping-dash bug-spline bug-5191...
xcb-window& (rgb24): 1 crashed! - gl-surface-source
xcb-window& (rgb24): 98 failed - arc-looping-dash bug-spline bug-519...
xcb-render-0_0 (argb32): 1 crashed! - gl-surface-source
xcb-render-0_0 (argb32): 55 failed - bug-seams clip-disjoint-hatchin...
xcb-render-0_0 (rgb24): 1 crashed! - gl-surface-source
xcb-render-0_0 (rgb24): 66 failed - bug-seams clip-disjoint-hatching...
xcb-fallback (rgb24): 1 crashed! - gl-surface-source
xcb-fallback (rgb24): 50 failed - bug-seams clip-operator extended-b...

to:

365 Passed, 160 Failed [1 crashed, 8 expected], 29 Skipped
Preamble: 1 failed - create-from-png
image (argb32): 1 crashed! - gl-surface-source
image (argb32): 41 failed - clip-operator extended-blend-alpha exten...
image (rgb24): 1 crashed! - gl-surface-source
image (rgb24): 31 failed - extended-blend-mask extended-blend-alpha-...
xlib (argb32): 1 crashed! - gl-surface-source
xlib (argb32): 104 failed - aliasing arc-looping-dash bug-spline bug...
xlib (rgb24): 1 crashed! - gl-surface-source
xlib (rgb24): 99 failed - aliasing arc-looping-dash bug-spline bug-e...
xlib-window (rgb24): 1 crashed! - gl-surface-source
xlib-window (rgb24): 97 failed - aliasing arc-looping-dash bug-splin...
xlib-render-0_0 (rgb24): 1 crashed! - gl-surface-source
xlib-render-0_0 (rgb24): 59 failed - clip-operator clip-push-group c...
xlib-fallback (rgb24): 1 crashed! - gl-surface-source
xlib-fallback (rgb24): 50 failed - bug-seams extended-blend-mask ext...
xcb (argb32): 1 crashed! - gl-surface-source
xcb (argb32): 50 failed - bug-spline hatchings huge-radial operator-...
xcb (rgb24): 1 crashed! - gl-surface-source
xcb (rgb24): 45 failed - bug-spline hatchings huge-radial overlappin...
xcb-window (rgb24): 1 crashed! - gl-surface-source
xcb-window (rgb24): 44 failed - bug-spline hatchings huge-radial ove...
xcb-window& (rgb24): 1 crashed! - gl-surface-source
xcb-window& (rgb24): 44 failed - bug-spline hatchings huge-radial ov...
xcb-render-0_0 (argb32): 1 crashed! - gl-surface-source
xcb-render-0_0 (argb32): 55 failed - bug-seams clip-disjoint-hatchin...
xcb-render-0_0 (rgb24): 1 crashed! - gl-surface-source
xcb-render-0_0 (rgb24): 49 failed - bug-seams clip-disjoint-hatching...
xcb-fallback (rgb24): 1 crashed! - gl-surface-source
xcb-fallback (rgb24): 32 failed - bug-seams extended-blend-mask exte...
Comment 2 Uli Schlachter 2014-10-03 18:35:09 UTC
Isn't the bug here really in the code that compares the images?

You are blaming test/buffer-diff.c line 80 which is:

   if ((row_a[x] & mask) != (row_b[x] & mask))

However, this mask value is set by the caller to either 0xffffffff or 0x00ffffff, depending on if the image has an alpha channel. In other words, this code correctly ignores the unused byte in RGB24 images.

The bug really is in test/pdiff/pdiff.c, function pdiff_compare(). This function does NOT do something like "cairo_surface_get_content (surface_a) & CAIRO_CONTENT_ALPHA ?  0xffffffff : 0x00ffffff" and treats all images as ARGB32.

This functions uses _get_red() etc to extract colors from the image:

 177 static uint32_t
 178 _get_pixel (const uint32_t *data, int i)
 179 {
 180     return data[i];
 181 }
 182 
 183 static unsigned char
 184 _get_red (const uint32_t *data, int i)                                                                                                                           
 185 {
 186     uint32_t pixel;
 187     uint8_t alpha;
 188 
 189     pixel = _get_pixel (data, i);
 190     alpha = (pixel & 0xff000000) >> 24;
 191     if (alpha == 0)
 192         return 0;
 193     else
 194         return (((pixel & 0x00ff0000) >> 16) * 255 + alpha / 2) / alpha;
 195 }

So the real bug is in this code. No idea how it can be easily fixed. An ugly work around would be to copy the images in case they aren't format ARGB32 yet via cairo so that cairo makes sure the alpha channel is set to 0xff....?
Comment 3 Massimo 2014-10-04 04:41:07 UTC
(In reply to Uli Schlachter from comment #2)
> Isn't the bug here really in the code that compares the images?
> 
> You are blaming test/buffer-diff.c line 80 which is:
> 
>    if ((row_a[x] & mask) != (row_b[x] & mask))
> 

I blamed test/buffer-diff.c because when that condition is
TRUE, the maximum diff computed includes also the non-alpha
channel:

http://cgit.freedesktop.org/cairo/tree/test/buffer-diff.c#n85

So my first observation was reading a test log in which
two identical images (argb32 and rgb24) had reported two
max diff values compared to the same reference.

yes, there are also other bugs in test/pdiff
Comment 4 Uli Schlachter 2014-10-04 09:05:38 UTC
(In reply to Massimo from comment #3)
> (In reply to Uli Schlachter from comment #2)
> > Isn't the bug here really in the code that compares the images?
> > 
> > You are blaming test/buffer-diff.c line 80 which is:
> > 
> >    if ((row_a[x] & mask) != (row_b[x] & mask))
> > 
> 
> I blamed test/buffer-diff.c because when that condition is
> TRUE, the maximum diff computed includes also the non-alpha
> channel:
> 
> http://cgit.freedesktop.org/cairo/tree/test/buffer-diff.c#n85
> 
> So my first observation was reading a test log in which
> two identical images (argb32 and rgb24) had reported two
> max diff values compared to the same reference.

Ah, I missed that part. From a quick look, I'd fix that by changing these two lines:

		    int value_a = (row_a[x] >> (channel*8)) & 0xff;
		    int value_b = (row_b[x] >> (channel*8)) & 0xff;

into

		    int value_a = ((row_a[x] & mask) >> (channel*8)) & 0xff;

Oh and line 101 needs to be changed as well, something like "&& mask != 0x00ffffff". Here it seems simpler to pass a cairo_format_t or cairo_bool_t into buffer_diff_core() and calculate the mask based on that.

> yes, there are also other bugs in test/pdiff

I'd rather fix these problems in pdiff instead of somewhere in the code producing these images. In theory, these two code paths are independent from each other and pdiff could be used from something else than just the test suite.
Comment 5 Massimo 2014-10-04 10:17:04 UTC
(In reply to Uli Schlachter from comment #4)
> 
> I'd rather fix these problems in pdiff instead of somewhere in the code
> producing these images. In theory, these two code paths are independent from
> each other and pdiff could be used from something else than just the test
> suite.

Agreed, I opted to fix the producer because a failing test
compares an image with many references, but is produced
only once, so with many failures in the test-suite this
extra work is not repeated many times. 

OTOH to speed up the execution of the test suite probably it could be
possible to save with the reference also its md5sum, so that for
the exact comparison one computes the md5sum of the output and compares
it against many references with only one image loop.
Comment 6 GitLab Migration User 2018-08-25 13:29:04 UTC
-- GitLab Migration Automatic Message --

This bug has been migrated to freedesktop.org's GitLab instance and has been closed from further activity.

You can subscribe and participate further through the new bug through this link to our GitLab instance: https://gitlab.freedesktop.org/cairo/cairo/issues/25.


Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.