[OpenJDK 2D-Dev] [9] Review request for 8142966 Wrong cursor position in text components on HiDPI display
Alexandr Scherbatiy
alexandr.scherbatiy at oracle.com
Fri Mar 25 18:35:16 UTC 2016
On 3/24/2016 11:27 AM, Phil Race wrote:
>
> I am not sure developers on OS X have the same history of expectation
> that metrics will not be changed but I think they may find it harder to
> spot there anyway because the rendering on OS X does not use hinting
> but it is an issue everywhere else.
> So pointing to a similar change on OS X doesn't help.
> The question is if you take a "low-dpi" display and draw a variety of
> the same long
> strings before and after the fix, with a variety of fonts,
> is there any case where they measure any differently.
> I think there must be if you throw away the hinted advances.
> Note that the most important case is Windows, not Linux, and
> you need to look at T2K not just freetype.
I generated difference for all fonts from local graphics environment
with size 12 between jdk with and without the fix for "low-dpi" display.
Results are in file [1] with simple format:
------------------------
Font: java.awt.Font[family=Aharoni,name=Aharoni Bold,style=plain,size=12]
[upper text width: 305] [upper text width: 306]
[long text width: 553] [long text width: 554]
[char [A] = 8] [char [A] = 9]
Font: java.awt.Font[family=Andalus,name=Andalus,style=plain,size=12]
[lower text width: 222] [lower text width: 220]
[upper text width: 279] [upper text width: 275]
[long text width: 501] [long text width: 495]
[char [t] = 4] [char [t] = 3]
[char [B] = 7] [char [B] = 6]
[char [J] = 4] [char [J] = 3]
[char [R] = 7] [char [R] = 6]
------------------------
Where lower text width is width of the text "the quick brown fox
jumps over the lazy dog", upper text width is width of the same text
with upper case and long text is sum of these both texts.
166 fonts out of 362 have difference.
Links [2] and [3] contains some images with difference between drawn
text before and after the fix. The text on the top is text before the
fix, the text on the bottom is text after the fix and text between them
shows the difference by red color.
[1]
http://cr.openjdk.java.net/~alexsch/8142966/font-size-diff/font-size-diff_00.txt
[2] http://cr.openjdk.java.net/~alexsch/8142966/diff-images.00/low-case
[3] http://cr.openjdk.java.net/~alexsch/8142966/diff-images.00/upper-case
Thanks,
Alexandr.
>
> -phil.
>
> On 03/09/2016 04:43 AM, Alexander Scherbatiy wrote:
>>
>> Hello,
>>
>> Could you review the proposed fix?
>>
>> Thanks,
>> Alexandr.
>>
>> On 13/02/16 00:16, Alexander Scherbatiy wrote:
>>> On 09/02/16 18:56, Sergey Bylokhov wrote:
>>>> Also probably it will be possible to test this via the public api
>>>> only(using the mix of the graphics transform + font transform).
>>>>
>>>> On 09.02.16 17:47, Sergey Bylokhov wrote:
>>>>> Some additional information.
>>>>> The Swing calculates the size of the components and location of the
>>>>> cursor based on the text metrics. And these text metrics based on
>>>>> advance. The Swing do not know to what destination it will be
>>>>> painted,
>>>>> it calculates the size based on the non-scaled(non-retina)
>>>>> destination.
>>>>> The problem occurs when we paint such component to the hidpi screen,
>>>>> because we get small round errors per glyph -> this causes a visible
>>>>> issueы when the text string is long. As a solution on the macosx the
>>>>> round operation was done in the users space instead of dev space.
>>>>>
>>>>> For example before the fix the size 1.4px was rounded to 3 (1.4 * 2 =
>>>>> 2.8 -> 3), but after the fix it will be 2 (1.4-> 1 * 2 = 2). So the
>>>>> Swing will be able to calculate correct size/location without
>>>>> information of the device scale. Note that the 3px(in dev space) is a
>>>>> kind of fractional coordinate in the user space(1.5). And the
>>>>> Swing does
>>>>> not work properly when fractional metrics are used, because it use
>>>>> ints
>>>>> as a coordinates.
>>>>>
>>>>> Note that the fix should be applied only when fractional metrics
>>>>> is off,
>>>>> otherwise we should not use any rounding. I am not sure that the
>>>>> current
>>>>> fix take it into account.
>>> The proposed fix is applied only when the fractional metrics are off.
>>>>>
>>>>>
>>>>> On 08.02.16 22:14, Jim Graham wrote:
>>>>>> Isn't the problem there that we are returning an integer as the
>>>>>> advance?
>>>>>> Why aren't we returning 7.35 as a value instead of 8?
>>> 7.35 is returned when fractional metrics are on.
>>>>>>
>>>>>> Also, shouldn't 7.35 round to 7 and 14.7 round to 15?
>>> This happens because when fractional metrics are on the glyph
>>> linearAdvance value is returned (for current case
>>> ftglyph->linearHoriAdvance = 7.35).
>>> But when fractional metrics are off the glyph advance is returned
>>> (ftglyph->advance.x = 8).
>>> This calculation is done by freetype library.
>>>
>>> Thanks,
>>> Alexandr.
>>>
>>>>>>
>>>>>> ...jim
>>>>>>
>>>>>> On 2/6/2016 7:28 AM, Alexander Scherbatiy wrote:
>>>>>>> On 05/02/16 23:39, Phil Race wrote:
>>>>>>>> Two things strike me when I read this
>>>>>>>>
>>>>>>>> 1) Initial surprise at how deep into the font code it goes.
>>>>>>>> Perhaps there are alternatives.
>>>>>>>>
>>>>>>>> 2) That it changes to using the linear metrics for measuring
>>>>>>>> advance.
>>>>>>>> Regardless of (1) I do not think (2) is correct. I am fairly
>>>>>>>> sure this
>>>>>>>> will lead to changes in text advance. It seems like it must throw
>>>>>>>> away adjusted metrics as a result of glyph hinting.
>>>>>>>>
>>>>>>>> I don't know what the fix should be, since I have not looked at
>>>>>>>> the
>>>>>>>> problem top-down, I am just seeing the bottom-up proposed
>>>>>>>> solution.
>>>>>>>> So all I can say for now is that it needs to be at least somewhat
>>>>>>>> different.
>>>>>>>
>>>>>>> There was the same issue on Mac OS X which has been fixed in the
>>>>>>> similar way:
>>>>>>> 8013569 [macosx] JLabel preferred size incorrect on retina
>>>>>>> displays
>>>>>>> with non-default font size
>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8013569
>>>>>>> http://cr.openjdk.java.net/~serb/7190349/webrev.04/
>>>>>>>
>>>>>>> The problem is that in many case for UI scale 2 we return to
>>>>>>> a user
>>>>>>> an unscaled value.
>>>>>>> But a value for transformed glyph advance, rounded and
>>>>>>> descaled can
>>>>>>> differ from just rounded glyph advance.
>>>>>>>
>>>>>>> For example, font[dialog, plain, 12] char 'a':
>>>>>>> transform: 12, advance: 7.35, rounded advance: 8
>>>>>>> transform: 24, advance: 14.70 round advance: 14
>>>>>>>
>>>>>>> and 8 does not equal 14 / 2.
>>>>>>>
>>>>>>> The solution for Mac OS X was to get the glyph advance using
>>>>>>> only
>>>>>>> font transform, round it and then apply the dev transform:
>>>>>>>
>>>>>>> CGGlyphImages.m:
>>>>>>> 481 advance = CGSizeApplyAffineTransform(advance,
>>>>>>> strike->fFontTx);
>>>>>>> 482 if (!JRSFontStyleUsesFractionalMetrics(strike->fStyle)) {
>>>>>>> 483 advance.width = round(advance.width);
>>>>>>> 484 advance.height = round(advance.height);
>>>>>>> 485 }
>>>>>>> 486 advance = CGSizeApplyAffineTransform(advance,
>>>>>>> strike->fDevTx);
>>>>>>>
>>>>>>> Thanks,
>>>>>>> Alexandr.
>>>>>>>>
>>>>>>>> -phil.
>>>>>>>>
>>>>>>>> On 01/27/2016 01:26 PM, Alexander Scherbatiy wrote:
>>>>>>>>>
>>>>>>>>> Hello,
>>>>>>>>>
>>>>>>>>> Could you review the fix:
>>>>>>>>> bug: https://bugs.openjdk.java.net/browse/JDK-8142966
>>>>>>>>> webrev: http://cr.openjdk.java.net/~alexsch/8142966/webev.00/
>>>>>>>>>
>>>>>>>>> The proposed fix rounds a glyph advance first and then scales
>>>>>>>>> it if
>>>>>>>>> UI scales do not equal to one.
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>> Alexandr.
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>
>>
>
More information about the 2d-dev
mailing list