RADV Gets Another Optimization For Micro-Benchmarks

[phoronix]

Phoronix: RADV Gets Another Optimization For Micro-Benchmarks

David Airlie and Bas Nieuwenhuizen's work on the RADV open-source Vulkan driver is quite relentless. David has posted yet another patch working on further optimizing the performance of this unofficial Radeon Vulkan driver living within Mesa...

http://www.phoronix.com/scan.php?page=news_item&px=RADV-Stop-Extra-Vertex-State

[theriddick]

Would love to see some patches that help RADV thread the workload better between CPU threads, atm I think that is a area where NVIDIA is winning quite a bit on with their Vulkan driver. Some Vulkan CPU thread loading benchmark results would be nice ;-)

[stupotace]

Would love to see some patches that help RADV thread the workload better between CPU threads, atm I think that is a area where NVIDIA is winning quite a bit on with their Vulkan driver. Some Vulkan CPU thread loading benchmark results would be nice ;-)

Does that even make sense for a Vulkan driver? I thought one of the main ideas of Vulkan, being closer to the metal, was to give control of such things to the applications themselves.

[nuetzel]

Would love to see some patches that help RADV thread the workload better between CPU threads, atm I think that is a area where NVIDIA is winning quite a bit on with their Vulkan driver. Some Vulkan CPU thread loading benchmark results would be nice ;-)

Dave, Bas or Samuel should comment definitely on this, but I can tell you that on my Xeon X3470 (Nehalem) development/testing system (I'm the Phoronix reader...), 'performance' governor didn't offer much (nearly zero) improvements (drop of CPU usage or fps increase) with vulkan but very notable with OpenGL (gl threading and R600_DEBUG=sisched works GREAT).

Conclusion:
Michael should always (mostly) run with 'performance' (maybe with 'sisched', too).
2. New comparison shouldn't harm...