Reply With QuoteThere are a number of benchmarks for individual operations; the problem is that AFAIK none of them really map well onto what a user sees, particularly when running through a compositor.
Humans click things and judge performance based on the *range* of times it takes for the desired action to settle down on the screen, typically after having gone through the driver stack multiple times -- eg once to figure out what got clicked, once to react and draw stuff, once to composite the results, and maybe once more to make the composited screen visible.
I am not aware of any benchmarks which track & measure end-to-end performance like that, or which measure response time multiple times and present some kind of distribution (since that is what best matches a user's perception of "performance").
Originally Posted by airlied
