This problem is discovered by the nightly regression test on 2006-01-22 : the file "pdfwrite Testform.v1.0.2.pdf (pkmraw/300/noband)" appears different. Analyzing it so far, the old (baseline) raster appears incorrect. The related patch http://ghostscript.com/pipermail/gs-cvs/2006-January/006240.html simply increases the halftone cache size : -#define max_tile_bytes_LARGE 4096 +#define max_tile_bytes_LARGE 65536 Since the cahce size must have no effect on raster, I believe that I observe a Ghostscript bug, which is visible with a small halftone cache only.
Created attachment 1996 [details] Testform.v1.0.2.pdf.pdf.pkmraw.300.0.gz The old (incorrect) raster, copied from peeves:~regression/regression/raster
Created attachment 1997 [details] out6480.ppm.gz A new raster, created on peeves with the revision 6480. Here is how I created it : svn checkout --username leonardo -r 6480 http://svn.ghostscript.com:8080/ghostscript/trunk/gs/ gs cd gs ./autogen.sh make cd .. rename gs gs6480 gs6480/bin/gs -Igs6480/lib:/home/regression/fonts/ -dBATCH -dNOPAUSE -sDEVICE=pdfwrite -sOutputFile=out6480.pdf -c false 0 startjob pop -f - <~regression/comparefiles/Testform.v1.0.2.pdf gs6480/bin/gs -Igs6480/lib:/home/regression/fonts/ -r300 -dBATCH -dNOPAUSE -sDEVICE=pkmraw -sOutputFile=out6480.ppm out6480.pdf gzip out6480.ppm
The revision 6479 creates a wrong raster, same as Testform.v1.0.2.pdf.pdf.pkmraw.300.0.gz attached to here. Thus, we need to debug the revision 6479, and port the patch to the current HEAD.
I am not seeing any issues with this. Creating pkmraw from the pdfwrite output looks the same to me as creating pkmraw from the original file. Igor does discuss that the issue is from changing max_ht_cached_tiles_LARGE (which is currently set to 8192). I am going to pass this one off to Ray to get his comments.
There have been several fixes for the halftoning and cache handling, and I have tested this with cache sizes of 4096 and 256 and there are no problems.