WIP: RFC virgl: Asnychronous command stream submission and more frequent flushes
With asynchronous command stream submission the penalty for submitting the command stream becomes low, because we don't wait for the transfer of the data to kernel space, hence we can flush the commans stream more often which can improve host utilization with ceratin work loads.
Running a few benchmarks show the following changes
Host:
CPU: AMD Ryzen 9 5900X
GPU: AMD Radeon RX 6700 XT
VM: Qemu/Gentoo| Benchmark | before | after | change (%) |
|---|---|---|---|
| Openarena | 301 | 413 | 37.2 |
| Unigine Tropics | 164 | 233 | 42.1 |
| Unigine Sanctuary | 195 | 408 | 109.2 |
| Unigine Heaven | 163 | 163 | 0.0 |
| Unigine Valley | 160 | 159 | -0.6 |
| Xonotic | 250 | 270 | 8.0 |
| GLmark2 2021.08.30 | 4672 | 4517 | -3.3 |
| GpuTest Volplosion | 13369 | 14859 | 11.1 |
WIP because I have to run a few more tests
Edited by Gert Wollny