Another Batch Of RadeonSI Optimizations Land In Mesa 21.3
Marek Olsak and other AMD developers have been working on an assortment of RadeonSI optimizations over the past number of weeks to reduce CPU overhead during certain driver operations and other improvements. The work is more than 42 patches in total.
Merged on Friday was the second batch of these RadeonSI optimizations for what has been reviewed and approved so far from that larger series.
Among the work merged yesterday includes:
- Not using Next-Gen Geometry (NGG) pass-through when culling is possible for better performance.
- Enabling shader-based primitive culling with polygon mode.
- Reducing the frequency of switching geometry shader fast launch on/off due to a performance hit from the extra context rolls.
- Eliminating the primitive discard compute shader since it doesn't always work, only relevant for GFX9/Vega and older, and is considered too complicated.
Other patches to improve efficiency of the driver and performance optimizations. No performance numbers were shared as part of this merge request or the individual patches as largely an assortment of micro-optimizations, but as we've seen with the open-source AMD Linux graphics driver stack over the years, every little bit helps.
This and other recent RadeonSI optimizations will be found in Mesa 21.3 due for release around November.