That's what we do now.  It sounds like someone checked in new checksums
without their corresponding new images, though, so the tests pass even
though the nominally expected PNGs are wrong.

- Pam

On Mon, Jun 22, 2009 at 11:40 AM, Greg Spencer <[email protected]> wrote:

> How about just running the pixel comparison only when the checksums don't
> match?  Still not ideal, of course.
> -Greg.
>
>
> On Mon, Jun 22, 2009 at 11:30 AM, Ojan Vafai <[email protected]> wrote:
>
>> This isn't the best, but it would be easy to add a flag to
>> run-webkit-tests that told it to always do the pixel comparison even if the
>> checksums matched.
>> Ojan
>>
>>
>> On Mon, Jun 22, 2009 at 10:55 AM, Evan Martin <[email protected]> wrote:
>>
>>>
>>> Just so I'm not all negative, my suggestions after consulting with Tony:
>>>
>>> 1) Make Linux behavior match Windows, ignoring the recommendation in
>>> the comments below.
>>> 2) Rebaseline everything on Linux. :(
>>> 3) Now converting from a PNG file to expected output is easy on all
>>> three platforms:
>>>   convert input.png rgba:- | swizzle_rgba_to_bgra | md5sum
>>>
>>> (Not certain if Mac uses BGRA images, though.)
>>>
>>> On Mon, Jun 22, 2009 at 10:24 AM, Evan Martin<[email protected]> wrote:
>>> > Since I'm waiting for a build I sat down to implement this.
>>> >
>>> > But our image checksums are not checksums of the image files :(, but
>>> > rather checksums of the pixels stored in the image.
>>> >
>>> > And Tony points out that our image checksumming is completely insane:
>>> >
>>> > =====
>>> >  // Fix the alpha. The expected PNGs on Mac have an alpha channel, so
>>> we want
>>> >  // to keep it. On Windows, the alpha channel is wrong since text/form
>>> control
>>> >  // drawing may have erased it in a few places. So on Windows we force
>>> it to
>>> >  // opaque and also don't write the alpha channel for the reference.
>>> Linux
>>> >  // doesn't have the wrong alpha like Windows, but we ignore it anyway.
>>> > #if defined(OS_WIN)
>>> >  bool discard_transparency = true;
>>> >  device->makeOpaque(0, 0, src_bmp.width(), src_bmp.height());
>>> > #elif defined(OS_LINUX)
>>> >  bool discard_transparency = true;
>>> > #elif defined(OS_MACOSX)
>>> >  bool discard_transparency = false;
>>> > #endif
>>> >
>>> >  // Compute MD5 sum.  We should have done this before calling
>>> >  // device->makeOpaque on Windows.  Because we do it after the call,
>>> there are
>>> >  // some images that are the pixel identical on windows and other
>>> platforms
>>> >  // but have different MD5 sums.  At this point, rebaselining all the
>>> windows
>>> >  // tests is too much of a pain, so we just check in different
>>> baselines.
>>> > ====
>>> >
>>> > To be more clear, here's a table of the platforms and their behaviors.
>>> > O=opaque, T=transparent.
>>> > (Sorry for my ghetto proportionally-spaced table here.)
>>> >
>>> >            Win   Mac  Lin
>>> > cksum    O     T       T
>>> > png        O      T      O
>>> >
>>> > I conclude that on Linux, you cannot go from the PNG file back to the
>>> > checksum in the presence of alpha.
>>> >
>>> >
>>> > Just for fun I played around a bit with commands like:
>>> >   convert path/to/pngfile rgba:- | md5sum
>>> > and wasn't able to repro the checksums I'm seeing.
>>> >
>>> > It looks ok from
>>> >   convert path/to/pngfile rgba:- | xxd -g4
>>> > (the RGBA<->BGRA problem doesn't apply for this black and while png
>>> file...).
>>> >
>>> > In summary: tears.
>>> >
>>> > On Mon, Jun 22, 2009 at 3:20 AM, Dean McNamee<[email protected]>
>>> wrote:
>>> >>
>>> >> Last week I updated our DEPS to pull in a newer version of Skia.  I
>>> >> was stumped at a few cases where the checked in PNG looked completely
>>> >> wrong, but yet it was passing on the buildbots.  There was no way that
>>> >> image could have been the output.
>>> >>
>>> >> It just dawned on me today, but I haven't verified it.  I can dig up
>>> >> my commit to verify it, but I'd say 99% sure this was the case.
>>> >>
>>> >> If the checksum is valid, we don't even go to the PNG.  Therefor I
>>> >> believe we have a bunch of layout tests where the checked in PNG is
>>> >> completely wrong, but the checksum is right.
>>> >>
>>> >> I don't have the time right now, but it would be great if someone
>>> >> could write a script and clean this up.
>>> >>
>>> >> Thanks
>>> >> -- dean
>>> >>
>>> >> >>
>>> >>
>>> >
>>>
>>>
>>>
>>
>>
>>
>
> >
>

--~--~---------~--~----~------------~-------~--~----~
Chromium Developers mailing list: [email protected] 
View archives, change email options, or unsubscribe: 
    http://groups.google.com/group/chromium-dev
-~----------~----~----~----~------~----~------~--~---

Reply via email to