Marek Squeezes More Performance Out Of RadeonSI In CPU-Bound Scenarios
The patch series is a set of command submission optimizations aimed to help trivial CPU-bound benchmarks to varying extents. In the very trivial glxgears, the patch series is able to improve the maximum frame-rates by around 10%.
The patches primarily clean-up the command submission code while one of the patches adds a new chunk type to the CS ioctl, which also require an updated libdrm.
These patches can for now be found on Mesa-dev.
Separately, Marek also sent out another patch for reducing LDS stalls by 40% for tessellation in RadeonSI.
With RadeonSI already performing competitive with the NVIDIA hardware/drivers on Linux and long ago having largely surpassed the AMD "PRO" OpenGL closed-source driver, it will be interesting to see how much more performance they can still squeeze out of AMDGPU+RadeonSI. On the feature front, they still need to get OpenGL 4.6 checked off the list but unfortunately it doesn't look like that big undertaking will land in time for the imminent Mesa 18.2 branching.