[OpenJDK 2D-Dev]  RFR JDK-8180055: Upgrade the Marlin renderer in Java2D
Prahalad Kumar Narayanan
prahalad.kumar.narayanan at oracle.com
Fri May 12 03:14:33 UTC 2017
Yes.. I wished to know the performance improvements between the versions of Marlin Renderer 0.7.5 and 0.7.4.
Thanks for the explanation. It helps.
Have a good day
From: Laurent Bourgès [mailto:bourges.laurent at gmail.com]
Sent: Thursday, May 11, 2017 5:40 PM
To: Prahalad Kumar Narayanan
Cc: Philip Race; Jim Graham; 2d-dev at openjdk.java.net
Subject: Re: [OpenJDK 2D-Dev]  RFR JDK-8180055: Upgrade the Marlin renderer in Java2D
2017-05-11 11:46 GMT+02:00 Prahalad Kumar Narayanan <HYPERLINK "mailto:prahalad.kumar.narayanan at oracle.com"prahalad.kumar.narayanan at oracle.com>:
+1 for the code changes.
Without going through the specifics, I imported the code & checked for successful build.
The build succeeded with the patch from your latest webrev link.
Thanks for the review.
Just for knowledge: You would have checked for the performance improvement with your improved Marlin renderer.
Can you share your insights and any specific use-cases that have benefited ?
Could you be more precise ? Do you mean the performance improvements between Marlin 0.7.5 and 0.7.4 ?
or in general by the Marlin renderer vs Pisces ...
In Marlin 0.7.5, I modified several aspects:
- double-precision (double) vs single-precision (float): ~ equivalent performance
- tile fills optimization: it benefits to large shapes (almost empty / full) like large circles or disks: up to 10% faster for large shapes measured with my MapBench tool (ellipse-fill commands from radius = 1 to 1000)
- higher precision for curve / quad approximation: it impacts performance as more segments are generated, sorted and rasterized ~ 10% globally in my MapBench runs.
I could provide some results if you want.
Finally I am going to work again on Java2D pipeline optimizations: avoid Path2D iterators, TexturePaint / CompositePaint (array / raster cache + loops)...
PS: I still plan one day to implement another approach to compute pixel coverage (exact) like agg / libart that could improve both quality and performance a bit (1 single scanline with more maths) but no more 8 scanlines.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the 2d-dev