On 28/02/2022 16:01, Rob Clark wrote:
On Mon, Feb 28, 2022 at 6:33 AM Tvrtko Ursulin
<tvrtko.ursulin@xxxxxxxxxxxxxxx> wrote:
On 25/02/2022 22:14, Rob Clark wrote:
On Fri, Feb 25, 2022 at 12:25 PM Rob Clark <robdclark@xxxxxxxxx> wrote:
From: Rob Clark <robdclark@xxxxxxxxxxxx>
Similar to AMD commit
874442541133 ("drm/amdgpu: Add show_fdinfo() interface"), using the
infrastructure added in previous patches, we add basic client info
and GPU engine utilisation for msm.
Example output:
# cat /proc/`pgrep glmark2`/fdinfo/6
pos: 0
flags: 02400002
mnt_id: 21
ino: 162
drm-driver: msm
drm-client-id: 7
drm-engine-gpu: 1734371319 ns
drm-cycles-gpu: 1153645024
Nice, so my vendor agnostic actually worked (with that single fixup of
accounting for the fact pdev tag is optional)?
Note that it might be useful to have a standardized way to report # of
cycles and max freq, so userspace tool can derive %utilization in
addition to just %busy
How do you define %utilisation vs %busy - I don't exactly follow since I
see the two as same?
so, say you are running at 50% of max clk, and gpu is busy 70% of the
time. The utilization is only 35% because the gpu could scale up the
clk to get more work done.
Got it, thanks. I don't think we have the equivalent on the i915 side
(we do have global frequency reporting via perf/PMU). In general things
like these I imagined would be defined as driver specific tags. If you
look at the drm-usage-stats.rst in my series, there is a "Driver
specific implementations" section in there which links to the i915 doc
section.
I've also put a "big fat" comment into the i915 fdinfo fops vfunc
pointing back to drm-usage-stats.rst which I think is useful when large
teams work on a driver. Not sure if that applies to msm so just mentioning.
Since this all works for you, would you mind applying your ack against
20220222140422.1121163-9-tvrtko.ursulin@xxxxxxxxxxxxxxx?
I need to get some updates r-b's for my series and then I submit it
again to Dave and Daniel for final acks.
After that, for a 2nd/follow-up phase, I plan to re-surrect the amdgpu
patch I had to make it compliant to common format, plus document the
option of engine utilisation tag being in percentage units as exposed
from that driver. And extending gputop to support that as well.
Regards,
Tvrtko
Looking at your patch I guess I don't understand the difference between
'elapsed' and 'cycles' inside your retire_submit(). Both are scoped to a
single context and are not global? If 'elapsed' is time context has
spent on the GPU, cycles isn't the same just in a different unit?
Correct, we capture (from GPU cmdstream) two counters both before and
after a submit (aka execbuf) runs, one is a fixed-rate counter, which
gives us elapsed time. The second is a counter that increments every
clk cycle, which gives us the # of cycles. With the two values, we
can calculate GPU frequency.
BR,
-R
Regards,
Tvrtko
