[SOLVED]libva-intel-media-driver now compiles much slower

Havin_it · Last edited by Havin_it on Mon Jul 17, 2023 8:38 pm; edited 1 time in total

Hi,

I noticed something odd happened a while ago with emerging libva-intel-media-driver, and would be grateful if anyone can work out what the cause is.

I track my build-times with a script that reads emerge.log, whose output is below:

stefan11111 · l33t Joined: 29 Jan 2023 Posts: 953 Location: Romania

Havin_it · Posted: Sun Jun 25, 2023 6:51 pm Post subject:

Probably not, this blob of flags is carried over from like 4 previous machines so there is some cruft no doubt, but I doubt they do any harm, no?

Maybe one day I will get around to tidying them up; but here I am taking my first opportunity to write a question about this bug/issue that costs me >2hrs of life and a bunch of electricity every month or two, look how long it took me to manage that and you get the idea of how much free time I have to play with vs. how I must prioritise things a little :wink:

ali3nx · Posted: Sun Jun 25, 2023 7:21 pm Post subject:

Josef.95 · Advocate Joined: 03 Sep 2007 Posts: 4695 Location: Germany

I noticed the new long built time here too:

logrusx · Advocate Joined: 22 Feb 2018 Posts: 2744

Havin_it · Posted: Mon Jun 26, 2023 8:32 am Post subject:

Hi ali3nx and thanks for the reply

sam_ · Developer Joined: 14 Aug 2020 Posts: 2161

ali3nx is definitely right and this is a fair point if someone is generally concerned about build times, but your point seems to be that you're wondering specifically what caused such a sudden large regression in this single package - which is a fair question. I'm now wondering the same..

The best way to figure this out, in lieu of any specific knowledge of this package, is to try build it manually from upstream (don't install it) and bisect it.

Havin_it · Posted: Mon Jun 26, 2023 8:41 am Post subject:

Thanks for replying Josef.95

Hu · Administrator Joined: 06 Mar 2007 Posts: 23123

I would look at it not as adding 20 minutes of compile time, but as multiplying the compile time by 6. Assuming this multiplier is approximately correct for other users, that will be especially significant for anyone using constrained hardware where the "good" times were much longer than 5 minutes.

As for whether to raise an issue upstream, it depends on the project. If the project routinely provides end-user support, you could try there. However, my preference would be to follow sam_'s advice and try to find more specifically why it is slow. Reporting the slowness to upstream will likely be needed eventually, but you may get a better reception if you can identify a specific commit as the culprit. It would be even better if your report can provide some informed commentary on whether it is reasonable. Hypothetically, if the offending commit is slow because broken dependency data is causing it to be single-threaded when it does not need to be, that would be clearly a bug. On the other hand, if the offending commit is slow because it introduced some notable new step, improving performance will be harder, because the step probably cannot be reverted.

ali3nx · Posted: Mon Jun 26, 2023 4:20 pm Post subject:

Havin_it · Posted: Tue Jun 27, 2023 9:42 am Post subject:

I've downloaded the 22.6.3 and .4 tarballs from github and plan to do some exploring with kdiff3 as soon as I'm able. Since I'm not a C++ speaker the odds are probably not great that I'll know it when I see it, but might as well have a look.

I do note from the build logs that the latter one compiled 16 more files than the former, so maybe in one of these the answer is found. We'll see.

I doubt I can do proper bisecting/rebuilding intensive approach any time soon. (Would have to learn how for a start.) If it came to that I would see more sense in raising a ticket just as an "FYI" if nothing else, since for all we know this may be something of whose effect the person who committed it is already well aware.

dmpogo · Advocate Joined: 02 Sep 2004 Posts: 3474 Location: Canada

Did compiler change around January on your machine ?

Havin_it · Posted: Tue Jun 27, 2023 7:37 pm Post subject:

Havin_it · Posted: Tue Jun 27, 2023 7:43 pm Post subject:

I went back to diffing the build logs to look at the actual compiler invocations, and compared the first really longwinded one, which is [18/1485] in 22.6.3 and [18/1501] in 22.6.4.

In the latter, there are the following added switches in the g++ commandline:

mattst88 · Developer Joined: 28 Oct 2004 Posts: 423

https://bugs.gentoo.org/910273

This is because of a change in sys-apps/sandbox.
_________________
My Wiki page

Havin_it · Posted: Mon Jul 17, 2023 8:38 pm Post subject:

sam_ · Developer Joined: 14 Aug 2020 Posts: 2161

(You can use 2.37 which has a workaround too.)

Havin_it · Posted: Thu Jul 20, 2023 4:05 pm Post subject:

mattst88 · Developer Joined: 28 Oct 2004 Posts: 423

Try building the same (i.e. older) version of libva-intel-media-driver and compare?
_________________
My Wiki page

Havin_it · Posted: Sat Jul 22, 2023 11:48 pm Post subject:

23.2.3 took ~43min with sandbox-2.37 (was ~30min with 2.29).

sam_ · Developer Joined: 14 Aug 2020 Posts: 2161

Havin_it · Posted: Sun Jul 23, 2023 12:54 am Post subject:

sam_ · Developer Joined: 14 Aug 2020 Posts: 2161

mattst88 · Developer Joined: 28 Oct 2004 Posts: 423

I'm not seeing any performance regression from sandbox-2.29 to sandbox-2.37 when compiling libva-intel-media-driver on a 64c/128t system: