[Mesa-dev] [PATCH 0/9] RadeonSI on AMDGPU: Command submission optimizations

[Marek Olšák]

https://pastebin.com/raw/itNDpPv0

Cleanups only: +10.85%
Total: +11.8%

Marek

On Fri, Jul 13, 2018 at 11:05 AM, Michel Dänzer <michel at daenzer.net> wrote:
> On 2018-07-12 07:26 AM, Marek Olšák wrote:
>>
>> The maximum glxgears FPS improves as follows.
>>
>> Initially: 13285
>> 8 patches: 14403 (+8.4% vs initial)
>> + patch 9: 15498 (+16.6% vs initial)
>
> Those are nice numbers, but I'm afraid they're almost meaningless
> without more information about how you obtained them:
>
>
> I ran glxgears three times on my development machine (Tonga, Ryzen 7)
> and let it print 10 lines of output each, see below. Observations:
>
> The maximum FPS varied between 8921 in the first run and 8519 in the
> second run, almost 5% difference.
>
> The average FPS varied between 8825 in the first run and 8390 in the
> second run, that's even slightly more than 5% difference.
>
> In summary, the variance I'm seeing between runs is almost as large as
> each of the two deltas in your numbers above.
>
>
> In order to get more confidence about the numbers, my suggestion would
> be to compare the average FPS instead of the maximum one (e.g. just
> adding up the numbers of frames from the first n lines of output of each
> run), and to do enough runs for the numbers to be statistically
> significant, e.g. using something like ministat (just put the total
> numbers of frames of each run on a separate line in a text file, one
> file for each test point, then feed the files to ministat, and it shows
> you the interval between the maximum/minimum difference between the test
> points at 95% confidence).
>
> (BTW, it's perfectly possible that this would result in even better
> numbers :)
>
>
> 1st run:
>
> 43508 frames in 5.0 seconds = 8701.432 FPS
> 44382 frames in 5.0 seconds = 8876.267 FPS
> 44473 frames in 5.0 seconds = 8894.457 FPS
> 44531 frames in 5.0 seconds = 8906.013 FPS
> 43774 frames in 5.0 seconds = 8754.668 FPS
> 42578 frames in 5.0 seconds = 8515.556 FPS
> 44379 frames in 5.0 seconds = 8875.649 FPS
> 44451 frames in 5.0 seconds = 8890.145 FPS
> 44572 frames in 5.0 seconds = 8914.330 FPS
> 44606 frames in 5.0 seconds = 8921.109 FPS
>
>
> 2nd run:
>
> 41095 frames in 5.0 seconds = 8218.930 FPS
> 42218 frames in 5.0 seconds = 8443.552 FPS
> 42561 frames in 5.0 seconds = 8512.142 FPS
> 42319 frames in 5.0 seconds = 8463.651 FPS
> 41450 frames in 5.0 seconds = 8289.875 FPS
> 41085 frames in 5.0 seconds = 8216.868 FPS
> 42593 frames in 5.0 seconds = 8518.580 FPS
> 42055 frames in 5.0 seconds = 8410.791 FPS
> 41651 frames in 5.0 seconds = 8330.172 FPS
> 42491 frames in 5.0 seconds = 8498.198 FPS
>
>
> 3rd run:
>
> 40701 frames in 5.0 seconds = 8140.039 FPS
> 41932 frames in 5.0 seconds = 8386.246 FPS
> 43277 frames in 5.0 seconds = 8655.284 FPS
> 44076 frames in 5.0 seconds = 8815.146 FPS
> 43535 frames in 5.0 seconds = 8706.958 FPS
> 42281 frames in 5.0 seconds = 8456.143 FPS
> 42347 frames in 5.0 seconds = 8469.265 FPS
> 40415 frames in 5.0 seconds = 8082.880 FPS
> 41142 frames in 5.0 seconds = 8228.342 FPS
> 40379 frames in 5.0 seconds = 8075.758 FPS
>
>
>
> --
> Earthling Michel Dänzer               |               http://www.amd.com
> Libre software enthusiast             |             Mesa and X developer