Intel Vulkan Driver Squeezes In A Few More Performance Optimizations For Mesa 23.1
Ahead of the Mesa 23.1 branching and feature freeze coming up in the next week or two, Intel's open-source graphics driver developers have been landing some last minute performance optimizations to benefit their "ANV" Vulkan driver.
First up, there is a fix for a compute state flushing not needed for Gen12.5 and newer. The CS stall flushing is only needed for pre-Gen12.5 graphics hardware where as up until a few days ago was also applied still to the latest Intel integrated/discrete graphics. By avoiding that flushing on Gen12.5+, games like Red Dead Redemption 2 enjoy a speed-up of around 3%.
Merged today for Mesa 23.1 were some uniform load optimizations to help reduce register pressure and in turn leading to more efficient register use in tested game titles like Assassin's Creed Odyssey, Cyberpunk 2077, Dota 2, Fallout 4, Witcher 3, Wolfenstein: Youngblood, and others.
Another optimization held up by the CS stall flushing change mentioned is indirect generated dispatch optimization but requires enabling via an environment variable or DriRC configuration file. Other recent ANV optimization work and other changes can be found via this Gitlab search. It's nice seeing some performance tuning work still land ahead of Mesa 23.1 on top of all their ongoing work around enabling the Xe kernel mode driver support.
First up, there is a fix for a compute state flushing not needed for Gen12.5 and newer. The CS stall flushing is only needed for pre-Gen12.5 graphics hardware where as up until a few days ago was also applied still to the latest Intel integrated/discrete graphics. By avoiding that flushing on Gen12.5+, games like Red Dead Redemption 2 enjoy a speed-up of around 3%.
Merged today for Mesa 23.1 were some uniform load optimizations to help reduce register pressure and in turn leading to more efficient register use in tested game titles like Assassin's Creed Odyssey, Cyberpunk 2077, Dota 2, Fallout 4, Witcher 3, Wolfenstein: Youngblood, and others.
Another optimization held up by the CS stall flushing change mentioned is indirect generated dispatch optimization but requires enabling via an environment variable or DriRC configuration file. Other recent ANV optimization work and other changes can be found via this Gitlab search. It's nice seeing some performance tuning work still land ahead of Mesa 23.1 on top of all their ongoing work around enabling the Xe kernel mode driver support.
2 Comments