RadeonSI Lands Yet Another Round Of Optimizations That Further Reduce CPU Overhead

Well known AMD open-source OpenGL driver developer Marek Olšák has landed another big batch of patches to further lower the driver overhead of this Linux OpenGL driver.

Marek merged two sets of optimizations today to Mesa 22.0-devel of micro-optimizations for this AMD Radeon OpenGL driver used on Linux for Radeon HD 7000 "Southern Islands" GCN 1.0 GPUs and newer up through the latest Radeon RX 6000 series RDNA2 graphics cards.

As with the many other RadeonSI optimizations this year, the focus is still principally on workstation OpenGL performance as measured by SPECViewPerf. The RadeonSI gaming performance has long been in excellent shape and far better than AMD's prior closed-source OpenGL driver that has shared sources with Windows. But with workstation applications still relying on OpenGL and that being an important area for AMD's business, RadeonSI optimizations aren't letting up.

These patches benefiting common Mesa and GLSL code will help applications that call glGetUniformLocation every frame.

This merge request also landing today provide various CPU overhead optimizations to the AMDGPU winsys code. This code in particular was found to particularly help SPECViewPerf 2020 with CATIA.

All of the RadeonSI optimizations in recent months are pushing the open-source OpenGL driver into uncharted territory for performance in becoming decisively better than the closed-source driver options. Too bad I don't have any Radeon Pro hardware for being able to deliver proper comparison benchmarks. We'll see what more lands in time for next quarter's Mesa 22.0 release.