Bug 692862

Summary:	Rounded boxes with zero height have a white triangle inside in gs 9.x
Product:	Ghostscript	Reporter:	Pavel Roskin <proski>
Component:	Graphics Library	Assignee:	Robin Watts <robin.watts>
Status:	RESOLVED FIXED
Severity:	normal	CC:	wl
Priority:	P4
Version:	master
Hardware:	PC
OS:	Linux
Customer:		Word Size:	---
Attachments:	Lilypond output Rendering to pngalpha - with white triangle Rendering to pnggray - without white triangle Patch to fix the white triangle

Description Pavel Roskin 2012-02-16 22:12:36 UTC

Created attachment 8360 [details]
Lilypond output

The bug was originally filed against Lilypond:
http://code.google.com/p/lilypond/issues/detail?id=2328

Lilypond wants to draw a line with rounded ends.  If rendering to PNG format, a write triangular area appears inside the line.

I found by bisecting ghostscript repository that the issue was introduced in commit 43b14b24fcc13d816dd41ca335d52cd3074bd0d8, between 8.71 and 9.01.

Reverting the patch fixes rendering.

Rendering to the pnggray device doesn't make the white triange, so ghostscript is inconsistent between devices.

Comment 1 Pavel Roskin 2012-02-16 22:14:22 UTC

Created attachment 8361 [details]
Rendering to pngalpha - with white triangle

Rendered by
gs -r300 -dBATCH -dNOPAUSE -sDEVICE=pngalpha -sOutputFile=box-alpha.png box.ps

Comment 2 Pavel Roskin 2012-02-16 22:15:11 UTC

Created attachment 8362 [details]
Rendering to pnggray - without white triangle

Rendered by
gs -r300 -dBATCH -dNOPAUSE -sDEVICE=pnggray -sOutputFile=box-gray.png box.ps

Comment 3 Pavel Roskin 2012-02-16 22:16:55 UTC

Created attachment 8363 [details]
Patch to fix the white triangle

This is essentially a revert of commit

commit cbcc4ac3c4d74ef2ca915c13c6d35a8a930ae4ff
Author: Robin Watts <robin.watts@artifex.com>
Date:   Wed Apr 7 13:13:57 2010 +0000

    Fix for bug 691228, problems with wide strokes with round caps/joins with
    antialiasing enabled.
    
    2 small problems in the same code. In the antialiased case, we stroke by
    forming 2 paths (one up either side of the line) and then reverse one and
    concatenate it. This should happen whenever we meet a closepath, but due to
    a misplaced return, was being missed out in some round join cases.
    
    Also, when using a round join in the case where 2 lines meet 'in-line' we
    would inadvertently leave the point on the wrong side of the line.
    
    No changes in localcluster testing as this only affects antialiased results
    and that's not tested currently.
    
    git-svn-id: http://svn.ghostscript.com/ghostscript/trunk@11031 a1074d23-0009-0410-80fe-cf8c14f379e6

Comment 4 Robin Watts 2013-03-07 11:15:30 UTC

(In reply to comment #3)
> Created attachment 8363 [details]
> Patch to fix the white triangle
> 
> This is essentially a revert of commit

Simple reversion is clearly not the right fix, as that reintroduces all the problems that caused bug 691228 in the first place.

I have implemented a fix that makes the fix dependent on the relative direction of the vectors, and this seems to fix both cases.

commit 6961e71d7b6b9162f6b75fae60469175a4c8d43e
Author: Robin Watts <robin.watts@artifex.com>
Date:   Thu Mar 7 00:45:41 2013 +0000

    Bug 692862: Fix stroking problem with antialiasing

    The fix for bug 691228 causes problems in some cases; when joining
    two segment that are directly in line with the next, we add a
    circular line cap to ensure correct rendering. Which side of this
    end cap we end up on depends on which direction the line then
    continues in.

Thanks for the report!