Request for help/advice in investigation of one interesting "huge FPS regression" bug

[Henri Verbeet]

On 3 May 2012 07:17, Alexey Loukianov <mooroon2 at mail.ru> wrote:
> Trying to pinpoint the cause using oprofile produced no valuable results: it
> either me not able to use this wonderful profiler correctly or the issue is of
> such kind that isn't easily tracked by oprofile.
>
Personally I think perf is a bit nicer to use than oprofile.
Nevertheless, two issues you're likely to run into are on the one hand
that the nvidia drivers don't have any kind of useful debugging
information, so you'll have a hard time profiling any time spent in
the driver, and on the other hand time spent waiting for the GPU
typically won't show up in the profile. This would happen for example
if we tried to upload to a vertex buffer that the GPU is still drawing
from. You can sometimes get useful information by measuring time for
some more high level operations like draws, blits, clears, etc. by
adding extra code to wined3d. In general tracking these kinds of
things down is just a lot of work and fairly hard though.