Hi Guillaume, To your question on IRC:
[14:37:37] <guillaume_g> dirk: is there a mean to be scheduled only on one type of machine (e.g. TX1 or TX2) in OBS? Maybe there is a flag or something? [14:40:41] <guillaume_g> it would be for next VPP version, which perform some runtime checks at configure and if scheduled on TX1 (and maybe HiSi) it would not run on other machines. And if it is build on TX2 (or others), it would not be efficient on TX1.
That sounds like we'd (short term) need to patch the configure script to explicitly be tuned for a particular target platform. Then build the binary n times - one for each optimization target. The long term fix is to work with OSEC for example and make sure that this becomes a real runtime check via IFUNCs for example, nothing checked at compile time. The same binary ideally runs fast everywhere then. Alex -- To unsubscribe, e-mail: opensuse-arm+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-arm+owner@opensuse.org
Hi Alex, ----- Alexander Graf <agraf@suse.de> a écrit :
Hi Guillaume,
To your question on IRC:
[14:37:37] <guillaume_g> dirk: is there a mean to be scheduled only on one type of machine (e.g. TX1 or TX2) in OBS? Maybe there is a flag or something? [14:40:41] <guillaume_g> it would be for next VPP version, which perform some runtime checks at configure and if scheduled on TX1 (and maybe HiSi) it would not run on other machines. And if it is build on TX2 (or others), it would not be efficient on TX1.
That sounds like we'd (short term) need to patch the configure script to explicitly be tuned for a particular target platform. Then build the binary n times - one for each optimization target.
Yes, that is my plan. I am checking internally with VPP people to check what would be the best way to patch and upstream it.
The long term fix is to work with OSEC for example and make sure that this becomes a real runtime check via IFUNCs for example, nothing checked at compile time. The same binary ideally runs fast everywhere then.
Some patches landed upstream recently for runtime detection, but not for this particular option apparently. Thanks, Guillaume
Alex
-- To unsubscribe, e-mail: opensuse-arm+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-arm+owner@opensuse.org
Am Mi., 2. Jan. 2019 um 09:24 Uhr schrieb Guillaume GARDET <guillaume.gardet@free.fr>:
explicitly be tuned for a particular target platform. Then build the binary n times - one for each optimization target.
That doesn't sound very appealing to me.
The long term fix is to work with OSEC for example and make sure that this becomes a real runtime check via IFUNCs for example, nothing checked at compile time. The same binary ideally runs fast everywhere then. Some patches landed upstream recently for runtime detection, but not for this particular option apparently.
That sounds like the right choice, until then we need to pick the lowest common platform features and optimize for that. Greetings, Dirk -- To unsubscribe, e-mail: opensuse-arm+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-arm+owner@opensuse.org
participants (3)
-
Alexander Graf -
Dirk Müller -
Guillaume GARDET