With the most recent Mesa 23.2 code as of Friday there may be now a fairly important efficiency optimization for Intel’s graphics driver stack that actually helps out Intel Arc Graphics DG2/Alchemist together with upcoming Meteor Lake graphics. Counter-Strike: International Offensive, for instance, was discovered to be 11% quicker now with this single driver change and different Vulkan apps/video games benefiting as properly.
The change that was merged yesterday allows L3 partial write merging for compressible surfaces and different appropriate circumstances for DG2 and shifting ahead with platforms like Meteor Lake. In flip this alteration of only a few strains of code is resulting in measurable speed-ups for Intel’s Arc Graphics discrete GPUs and serving to to slim the hole between the Home windows and Linux graphics driver efficiency.
Longtime Intel open-source Linux graphics driver engineer Francisco Jerez defined within the merge request:
“This permits L3 partial write merging for a lot of circumstances that appear to be getting by chance disabled by the kernel, which was inflicting a severe efficiency bottleneck on DG2 and MTL platforms. The “Compressible Partial Write Merge Allow”, “Coherent Partial Write Merge Allow” and “Cross-Tile Partial Write Merge Allow” bits in L3SQCREG5 have been anticipated to be enabled by default (and confusingly, they even learn off as enabled should you ran ‘intel_reg learn 0xb158’ on an idle system), however they’re getting clobbered throughout 3D context initialization by an i915 workaround.
Enabling L3 partial write merging of compressible surfaces particularly appears to extend rendering fillrate by over 3x in some circumstances (e.g. the “VulkanFillRate/FillRateGPU/decision:1[0-3]/format:*/mix:0” fillrate-bound microbenchmarks). Vital enhancements will also be reproduced in most real-world workloads we have examined thus far, e.g. Counter Strike GO improves by ~11%, Shadow Of the Tomb Raider improves by ~5.5%, and AztecRuins-VK improves by ~6.5% on DG2-512 — Thanks lots to Caleb Callaway for these figures. No regressions have been noticed thus far.
Despite the fact that this patch may strike as surprisingly easy for such a big payoff, it is the results of @fjdegroo and I attempting to root-cause the rendering efficiency hole of DG2 on Linux vs Home windows on and off over the past 12 months, and a few of the OA statistics captured by Felix early this month have been drastically useful for me to attach the previous few dots, so Felix deserves a giant chunk of the credit score for this work.”
Seeing measurable wins for varied Linux video games is thrilling and for sure I will be firing up some Arc Graphics A750/A770 Linux benchmarks this weekend to take a look at the advantages of this alteration throughout extra video games on Linux and Steam Play titles.
This alteration is presently within the Mesa 23.2-devel Git code, which can be out as steady round late August or September, whereas it stays to be seen but if the few strains of code are deemed cheap sufficient for back-porting to the present Mesa 23.1 steady sequence in a future level launch.
