[gs-cvs] rev 8880 - trunk/gs/lib

[leonardo at ghostscript.com]

Author: leonardo
Date: 2008-07-27 02:29:37 -0700 (Sun, 27 Jul 2008)
New Revision: 8880

Modified:
   trunk/gs/lib/pdf_draw.ps
Log:
Fix (PDF interpreter) : A temporary work around huge images with a soft mask.

DETAILS :

Bug 689080 "PDF Crash with Transparency at 2880 dpi".

 
This is a temporary workaraound for the problem,
see comment in code.

With this patch the test case doesn't crash,
but it needs  -dBufferSpace=500000000 to complete in 15 minutes
on Core 2 Duo 2.4 MHz with debug build.
The test case is pretty big anyway - it generates a 4.2 gigapixel raster.

Thankls to Ray for obtaining the buffer space from a device.

I think the old code is incorrect because it looks applying the
mask two times : first when installing a mask buffer with
.begintransparencymaskimage in doimagesmask, and second time when
processing a Type 103 image in C code. The raster difference
is likely caused by that. I don't perform a deeper analyzis,
because it's a temporary fix which needs to be replaced soon.
All that is written in this paragraph are not sure,
because there is no time for a deeper analyzis before 8.63 release.
When we'll complete Type 103 images (with rasterizing after clist
and with a proper handling in pdfwrite), .begintransparencymaskimage
will likely go away.

EXPECTED DIFFERENCES :

Bug689369.pdf (a minor raster difference at the bottom of nthe page).


Modified: trunk/gs/lib/pdf_draw.ps
===================================================================
--- trunk/gs/lib/pdf_draw.ps	2008-07-27 07:03:42 UTC (rev 8879)
+++ trunk/gs/lib/pdf_draw.ps	2008-07-27 09:29:37 UTC (rev 8880)
@@ -1119,10 +1119,37 @@
     doimage
   } ifelse
 } bdef
+
+/is_big_mask {  % <mask> is_big_mask <bool>
+  1 0 dtransform dup mul exch dup mul sqrt add
+  0 1 dtransform dup mul exch dup mul sqrt add mul
+  currentdevice getdeviceprops .dicttomark /BufferSpace .knownget not {
+    4000000 % hack: Can't get the real default value from C source.
+  } if
+  2 % arbitrary
+  div gt
+} bind def
+
 /doimage {	% <imagemask> doimage -
 		% imagedict is currentdict, gets popped from dstack
   DataSource exch
-  PDFusingtransparency {
+
+  PDFusingtransparency
+  currentdevice .devicename /pdfwrite ne { 
+    % This is a temporary workaround for the bug 689080,
+    % which is done under a rush of 8.63 release.
+    % The purpose is to disable a conversion of an image with soft mask
+    % into a Type 103 image, which currently allocates a full mask buffer
+    % before writing clist. Note that imagemask installs
+    % the transparency mask anyway (likely another bug), 
+    % so the image is still masked.
+    % With this workaround the Matte color is not working (ignored).
+    is_big_mask not 
+  } {
+    true % pdfwrite doesn't need the workaround explained above,
+         % and cannot work with it. 
+  } ifelse 
+  and {
     currentdict /SMask knownoget
   } {
     false