HD Graphics 620: VAAPI performs poorly

"Recently" I've stumbled over the temperature issues described in #953 (closed) (which is likely a duplicate of #614 (closed)) and made in addition the observation that (even under kernel 5.2, where, AFAIU, the issue of #614 (closed) shouldn't be present) playing back videos with VAAPI result sometimes in higher temperatures than playback without hardware acceleration.

@ickle suggested to file a separate bug for this.

And in general, it seems to perform pretty poor, playing back e.g. a 1920x1080 H264 video in fullscreen on a HiDPI display, causes the CPU/GPU reach 100°C ... but maybe the HD620 isn't just powerful enough... still it shouldn't perform worse than pure software acceleration, as it does in some cases at least.

For details of my testing methods, versions, etc. please have a look at #953 (comment 380857) .

Maybe someone could have a look :-)

Thanks, Chris.

mentioned in issue #953 (closed)

Oh, it should be able to do 1080p decode at minimum frequency. (The capacity is something like 10 1080p streams.)

Could you please record the power usage (those perf commands) for both decode methods? Hmm, in fact, lets measure the gpu usage as well, so put this into a script (hopefully the glob works!)

perf stat -a -x, -r 1 \
        -e "power/energy-pkg/" \
        -e "power/energy-cores/" \
        -e "power/energy-gpu/" \
        -e "i915/actual-frequency/" \
        -e "i915/rc6-residency/" \
        -e "i915/rcs0-busy/" \
        -e "i915/bcs0-busy/" \
        -e "i915/vcs*-busy/" \
        $*

(./myperf.sh mpv...)

I assume this should read -e "i915/vcs0-busy/"?

Also on which kernel (5.2 or 5.4 without the RC6 fix) and under which DE should I run it?

I was hoping the globbing worked, if not, just "i915/vcs0-busy/"

As for setup, I don't mind -- just record it so we can see the differences in environment. Usually use the latest stable commits and we'll go from there. (With the other end point drm-tip, which does need either a timestamp or commit id.)

added Community label

@calestyo did you get chance to generate the perf results as suggested in earlier comments?

Please provide the command that is being used for playing the videos with and without VAAPI. Also please share the GPU usage results using intel_gpu_top (part of IGT tools).

added platform: KBL label

I'm waiting for 5.5 to land in Debian, since it's probably pointless to do these tests on the outdated 5.2.

@calestyo Sure. :)

If the issue still persist, please provide few other details of the bug (apart from the details provided in bug #953 (comment 380857)).

1. The details asked in comment #956 (comment 396796)
2. How was the performance measurement done, as in VAAPI performance was weak?
3. Please confirm if the videos alone were played during your measurements, i.e., were there any other application running in background. The temperature measured while running the same kind of videos simultaneously did not show any discrepancy as mentioned.

added UX label

Well the issue is still present but from my side I guess you can close the issue. Since this project is not able/willing to fix a complete showstopper regression like 614 I'll simply have to switch to different hardware and sell the current notebook as oven or heating on ebay.

#614 (closed) is completely and utterly fixed.

closed

No, this is not true... unfixed and still fully and utterly breaking all affected systems.

Just confirmed, still the case in 5.5.13.

Please provide the details requested in comment #956 (comment 400161).

First of all, you may want to notice my testing updates in #953 (comment 457907)

Second I've just tried to do the: perf stat -a -x, -r 1 -e "power/energy-pkg/" -e "power/energy-cores/" -e "power/energy-gpu/" -e "i915/actual-frequency/" -e "i915/rc6-residency/" -e "i915/rcs0-busy/" -e "i915/bcs0-busy/" -e "i915/vcs0*-busy/" $*

as asked by Chris Wilson above... that fails already now:

# perf stat -a -x, -r 1 -e "power/energy-pkg/" -e "power/energy-cores/" -e "power/energy-gpu/" -e "i915/actual-frequency/" -e "i915/rc6-residency/" -e "i915/rcs0-busy/" -e "i915/bcs0-busy/" -e "i915/vcs0*-busy/" $*
event syntax error: 'i915/vcs0*-busy/'
                          \___ unknown term

valid terms: i915_eventid,config,config1,config2,name,period,percore

Initial error:
event syntax error: 'i915/vcs0*-busy/'
                          \___ unknown term

valid terms: i915_eventid,config,config1,config2,name,period,percore
Run 'perf list' for a list of valid events

 Usage: perf stat [<options>] [<command>]

    -e, --event <event>   event selector. use 'perf list' to list available events

This was under 5.5.13 with intel_pstate=disable and with gnome-shell-classic.

Any ideas what I'm doing wrong or what might have changed here? :-D

Thanks.

reopened

Saw my mistake o.O ... will do proper measurements tomorrow.

@banns @ickle I think I've added all the data you were asking for.

Took a bit longer because I've made some extensive set of tests (so far only for 5.5.13/intel_pstate enabled) which takes quite a while... and after I've completed the first set I found an error in my testing.

See #953 (comment 463451) for the details.

It should contain everything you wanted, the perf stat and intel_gpu_top (and more). Also, as for @banns questions:

2. the performance is weak with respect to the CPU temperatures going through the roof (i.e. generally hitting 100°C)

3. Confirmed that no other major programs were running(see detailed notes in the link above). Not sure what you mean exactly with "The temperature measured while running the same kind of videos simultaneously did not show any discrepancy as mentioned."

Cheers, Chris.

• "The temperature measured while running the same kind of videos simultaneously did not show any discrepancy as mentioned." *

@calestyo I tried to reproduce the issue based on the details provided in #953 (comment 380857), there was no spike in cpu temperature (as high as 100 degree) or performance degradation.

• "Confirmed that no other major programs were running"*

When you say no other major programs were running, what are the programs that are running in background? The reason for asking is, these said surges may also be caused due to some of the background running programs.

I tried to reproduce the issue based on the details provided in #953 (closed) (comment 380857), there was no spike in cpu temperature (as high as 100 degree) or performance degradation.

Well I'd guess the whole thing might show up only on certain CPU models and/or combinations with other hardware?! Otherwise many other people would probably also suffer from it (like it was the case with the RC6 bug).

The 100°C is really mostly an issue with playing back videos. For the other tests (moving windows in circles, unhide brute, SHA512 verifications, or e.g. just press&holding a key in e.g. Evolution) things go usually "just" into the 70-90°C or perhaps 95°C region.

Have you had a chance too look at the extensive tests I've posted at lkml (and which I've mentioned at: #953 (comment 457907) )?

From these, and the plots therein I'd have said that there is likely also something wrong which has nothing directly do to with the GPU... I mean the plots clearly show that e.g. unhide brute or the SHA512 verification runs much hotter with ips=on and additionally much hotter in 5.5 than 5.2 (with both ips on and off).

And have you had a chance to look at the further comments (#953 (comment 459076)) and specifically the test-series just with mpv and different parameters (https://github.com/calestyo/cpu-tests) there?

These pretty much show that even for the low res version (though still 60fps) I always hit the 100°C (in both, cinnamon and gnome shell clasic)... and @ickle mentioned above the GPU should be able to decode 10 streams of 1080p.

When you say no other major programs were running, what are the programs that are running in background? The reason for asking is, these said surges may also be caused due to some of the background running programs.

Attached is the ps ax output as it was under my tests (for cinnamon and g-s-c). ps-ax.cinnamon ps-ax.gnome-shell-classic

It looks like quite a number, but all these are mostly idling processes which effectively sleep all the time. And as I've said in my test report, I've always had cron/anacron disabled.

I've just seen that some systemd timers are active, but I've just checked it and given their settings, none should have run during the times at which I've tested. And even if,... all are rather short things, which either exit immediately (e.g. e2scrub_all.timer - I have no ext* fs) or wouldn't take very long. So at best these would produce some little spikes for a few seconds, but not explain the general picture that the plots show (and which matches my daily experience when actually using the laptop).

btw: I should perhaps add, that video playback (especially in fullscreen and with highres videos) always used to cause pretty high temperatures on that system (where I'd have expected that it should be barely noticeable at all)... i.e. even before 5.2.

I just got way much worse, after 5.2 (basically always hitting 100°C)

(haven't yet found the time to do further analysis, tests or plot... probably beginning next week)