HI Petr,
On 4/24/2020 11:25, Petr Štetiar wrote:
(adding atf-build's email to Cc: loop)
Zoltan HERPAI <wigy...@uid0.hu> [2020-04-23 17:24:04]:
Hi,
On 4/20/2020 16:29, Petr Štetiar wrote:
Update ATF to latest v2.3-rc1 version and use reproducible bl31 binaries
from atf-builds project.
atf-builds project started recently offering reproducible ATF binaries
which is more convenient then building very tiny ATF artifacts from
scratch during each build, thus saving considerable build resources. It
has reduced build from scratch time on my CI builder for sunxi/cortexa53
target from 57min down to 29min.
Run tested on a64-olinuxino-emmc.
NAK for now, sorry - they don't provide H6 builds which has a separate plat,
and the A64 builds can't be used on it.
I'm probably missing something here, where is this H6 used right now? Or you
plan to add this soon? I'm quite lost here as I did build/run test this
change, didn't noticed any breakage. I'm quite lost here.
Yes, H6 is in the works for some time - no breakage for now on this
change, but would break upcoming work.
If you can ping them to release H6 builds as well, then let's re-discuss.
You or anyone else can ping them[1] with a simple PR which would add other
binaries. I did the same with "Add sun50i_a64 target"[2] for a64.
That sounds great, if it's that easy to ask for new targets / binaries
to be built, I'll prep a PR for them to add sun50i_h6.
atf-builds started in "rockchip: add support for Pine64 RockPro64"[3],
otherwise we would need to build another toolchain (build time expensive) just
to get small ATF binary. Using reproducible ATF binaries seemed better
solution to me, then for example having Linux only solution (due to downloaded
Linux toolchain) in arm-trusted-firmware-mvebu.
Also, if you could share how the speed-up was measured only with this
change, that'd be great.
I'm building my staging tree for some time on GitLab CI, using same
steps/config as used on the project's buildbots. Most of the builds are around
30min, so those almost double build times for a53 are easily noticeable.
Thanks for the details. I don't see that this helps if you're building
images yourself for the Allwinner A53 boards as you need to compile the
toolchain anyway, but if this helps on the buildbots, if we can ask H6
images to be built, and everyone agrees that using pre-built ATF images
is the safe way forward, then I won't stop anyone from doing so.
https://gitlab.com/ynezz/openwrt/-/jobs/518186436 (29min, sunxi/cortexa53 with
atf-builds)
https://gitlab.com/ynezz/openwrt/-/jobs/502628921 (57min, sunxi/cortexa53)
https://gitlab.com/ynezz/openwrt/-/jobs/497486792 (56min, sunxi/cortexa53)
https://gitlab.com/ynezz/openwrt/-/jobs/496212640 (57min, sunxi/cortexa53)
https://gitlab.com/ynezz/openwrt/-/jobs/485014798 (27min, sunxi/cortexa8)
https://gitlab.com/ynezz/openwrt/-/jobs/485014797 (32min, sunxi/cortexa7)
1. https://github.com/atf-builds/atf#adding-new-build-targets
2. https://github.com/atf-builds/atf/pull/1
3. https://github.com/openwrt/openwrt/pull/2917#issuecomment-614329227
Thanks,
Zoltan H
_______________________________________________
openwrt-devel mailing list
openwrt-devel@lists.openwrt.org
https://lists.openwrt.org/mailman/listinfo/openwrt-devel
