Very poor performance in GL_ARB_compute_shader tests
zink-wip @ a7dc6178
AMD Radeon R9 Fury X (FIJI, GFX8)
When running GeeXLab_0.41.2.0_linux64.zip with d07-gl43-arb-compute-shaders-particles-simulation
demo, I got very bad performance with zink on radv:
name=geex-linux kind=normal driver=zink sub=gl4-arb-compute-particles opts= measurement=avg_fps unit=fps avg=19.6381 cssd=0.0239988 rcssd=0.122206% count=7 peak_vram=418.43 peak_rss=213.062 data= 19.6667 19.6333 19.6 19.6333 19.6333 19.6333 19.6667
name=geex-linux kind=normal driver=radeonsi sub=gl4-arb-compute-particles opts= measurement=avg_fps unit=fps avg=377.019 cssd=2.65366 rcssd=0.703852% count=7 peak_vram=382.93 peak_rss=223.523 data= 375.267 377.367 375.067 380 377.333 373.667 380.433
That is 5% performance only.
Downloads: https://www.geeks3d.com/dl/show/502 for GeeXLab and https://www.geeks3d.com/dl/show/50103 for GL4 demopack.
Run (example):
user@debian:~/Downloads/GeeXLab_linux64$ ~/zink-opt ./GeeXLab '/demofile="./demos/other/geexlab-demopack-gl4x/d07-gl43-arb-compute-shaders-particles-simulation/main.xml"' /width=1920 /height=1080 /benchmark /benchmark_duration_ms=30000 /nogui /no_menubar /benchmark_log_score /no_online_check
user@debian:~/Downloads/GeeXLab_linux64$ tail -1 _geexlab_scores.csv
(Note: The path to demofile must be relative, unless /demofile_abs_path=1
is also used).
The other ARB_compute_shader demos, for example ./demos/other/geexlab-demopack-gl4x/d08-gl43-arb-compute-shaders-particles-ssbo-grid/grid_deformer_v2.xml
, do behave much better, but still not great.
Also, the code somehow assumes reasonably small frametimes, which is tied to simulation timestep. So poor fps, would actually result in incorrect simulation in this benchmark. But same behaviour can be reproduced with radeonsi (i.e. by running multiple instances, or zink and radeonsi in parallel).