December 16, 2021 4:17 AM, "Henri Verbeet" hverbeet@gmail.com wrote:
Greatly improves performance in Control.
This doesn't tell someone reading the commit log all that much. What's "greatly"? Is that a 5% increase? Or perhaps 20%? 500%? Perhaps at least as important, what is the baseline? If it's running at 200 fps, and this takes of 2.5ms, that's going to double fps, yes; on the other hand, if it's running at 40 fps, and this takes of 10ms, that would be a fair bit more interesting, even if the percentage increase is a bit smaller. What hardware/driver/etc. is this on? Also, ideally that
IIRC on an RX 580 it was around 10% when fps dropped to ~50, so it saved about 2 ms. Perhaps not "greatly" depending your definition, but I took it verbatim from the original patch.
It's more complicated once optimised batched copies are added. Redundant copies are detected in the loop there, so this code path will handle much fewer descriptors which are probably unlikely to be redundant or make a noticeable performance impact.