<anholt_>
the plan is to do this for intel as well. maybe radv some day.
<zmike>
oh boy
<zmike>
can't wait to find out that my whitespace cleanup accidentally the whole perf
<anholt_>
limited by the quality of the single-frame tracing, of course. the plan for that is to move to having 6 frames that we loop, so that you don't have dependencies if the game is recycling buffers every 2 or 3 frames.
<zmike>
I'm still puzzled by the efficacy of it
<anholt_>
hm?
<zmike>
certainly it's useful as a reference for A against A over time (e.g., tracking zink perf against itself) but between drivers seems a bit more nebulous
<zmike>
...which is not to say that it's not useful, just that I'm confused
<anholt_>
do you feel that between-driver comparisons are not useful?
<anholt_>
(I'm not sure what you're getting at)
<zmike>
I'm not sure whether they are or are not
<zmike>
is what I'm saying
<zmike>
any cpu-intensive trace seems to be not super useful since the perf varies wildly from driver to driver
<zmike>
but more gpu-based stuff is a bit closer to actual benchmark territory
<anholt_>
the trace perf results are supposed to be tracking only gpu time
<anholt_>
(it uses time elapsed)
<zmike>
I've actually had some traces which don't work with time elapsed somehow
<anholt_>
which is dubious in its own way
<zmike>
which is another oddness
<zmike>
in any case, it'll be neat to have a dashboard up to look at