On Thu Apr 4 16:43:51 2024 +0000, Stefan Dösinger wrote:
I constructed a simple benchmark that just makes a bunch of draw calls
per frame (and nothing else). That benchmark is artificial, shows a consistent performance hit from this commit: Fyi, I have a collection of microbenchmarks here: https://gitlab.winehq.org/stefan/perftest Though most are stuck in d3d9 times. Once upon a time I was running them in an automated fashion daily...
Ah sorry, I wasn't clear enough.
The difference in Grounded vs not submitting happens because of the aforementioned GPU scheduling (and that's still true on the same machine with Mesa 24.0.4-1).
What I tested here, and was a bit too sleep-deprived to properly explain, was adding an *extra* few calls to QueryPerformanceCounter() in wined3d_context_vk_get_command_buffer(). I didn't see a difference between one call and five calls. I did see a difference with 100 calls, so it does matter, it's just not clear whether *one* is enough to matter.