Customer has a file with a moderate-sized jpeg (2000x2800) masked by a high resolution (10000x14000) JBIG2 image. Implementing the soft mask uses about 300 MB of ram. What optimizations can be done to reduce the footprint for this process, either downsampling or doing the composite operation in bands?
Created attachment 1881 [details] example file demonstrating the footprint issue
Acrobat can display the full page using only 80 MB of ram, but that may not be at the full 300 dpi.
To follow up, 8.53 has more intelligent banding for output devices derived from gdevprn, but I didn't see a huge difference with my own testing between HEAD and 8.51. Ray also suggests controlling the overall memory consumption with the -dBufferSpace= device parameter. I can however view the file with the X11 output device on linux at 72 or 100 dpi with < 100MB maximum allocation, and create a 300 dpi ppmraw with < 80 MB with both 8.51 and head.
80 Mb seems quite reasonable
Adjusting priority. The resolution for this is to port the smarter banding logic to non-prn devices, but we have no immediate plans to do so. Leaving this open since others are litle to trip over the same issue.
Assigning to Igor since this is similar to the 'big pattern' bug fix method that he is implementing using the clist (for bug 688396). Changing priority to 'enh' since the customer is happy for now and removing the customer ID since it isn't really a customer issue now. Was customer 581.
Returning to Support. We improved the transparency with banding. Please re-test with current revision. Is this a customer bug ? Which customer ?
Ghostscript processes this file using a reasonable amount of RAM with the current rev (8814). Improvements in the clist transparency code may have fixed this.