Mesa 24.1 Zink Lands "Super Fast" Merge Request To Optimize IO
Mike Blumenkrantz with Valve's open-source Linux graphics driver team has merged a big optimization / bug fixing effort he's recently been tackling for the Zink OpenGL-on-Vulkan driver.
Last week Blumenkrantz blogged about this Zink initiative in his usual colorful character. With that he concluded, "So far, all I’ve seen is about a 6000% improvement across the board. Finally, it’s over." Of course, that's not necessarily his scientific measurement but long story short he's been tackling a big project on Zink around IO optimizations.
On Friday that was followed up with this brief blog post:
The merge is the zink: compiler rewrites / opt_varyings for delivering some nice IO optimizations to Zink. No benchmark numbers were provided but I'll be giving the ZINK_DEBUG=ioopt mode a whirl soon for looking at the performance difference. As stated by Mike, the code should be enabled by default once a new OpenGL conformance test suite (CTS) release happens to fix broken tests in this mode.
Last week Blumenkrantz blogged about this Zink initiative in his usual colorful character. With that he concluded, "So far, all I’ve seen is about a 6000% improvement across the board. Finally, it’s over." Of course, that's not necessarily his scientific measurement but long story short he's been tackling a big project on Zink around IO optimizations.
On Friday that was followed up with this brief blog post:
Quick Post
Super Fast
Just a quick post to let everyone know that I have clicked merge on the vroom MR. Once it lands, you can test the added performance gains with ZINK_DEBUG=ioopt.
I’ll be enabling this by default in the next month or so once a new GL CTS release happens that fixes all the hundreds of broken tests which would otherwise regress. With that said, I’ve tested it on a number of games and benchmarks, and everything works as expected.
Have fun.
The merge is the zink: compiler rewrites / opt_varyings for delivering some nice IO optimizations to Zink. No benchmark numbers were provided but I'll be giving the ZINK_DEBUG=ioopt mode a whirl soon for looking at the performance difference. As stated by Mike, the code should be enabled by default once a new OpenGL conformance test suite (CTS) release happens to fix broken tests in this mode.
21 Comments