09:06 <github> [hpx] StellarBot pushed 1 new commit to gh-pages: https://git.io/fNm0G

09:06 <github> hpx/gh-pages 4a85a3c StellarBot: Updating docs

11:17 <jakub_golinowski> Is it possible to "call" hpx:dump-config from the ini file?

12:39 <jakub_golinowski> M-ms, hkaiser: as asked above: is it possible to somehow enforce dumping hpx config without being able to pass the CL --hpx:dump-config to the runtime?

12:40 <hkaiser> jaafar: no, but why are you not able to pass it to the runtime?

12:40 <hkaiser> jakub_golinowski: ^^

12:40 <hkaiser> sorry jaafar /

12:42 <jakub_golinowski> hkaiser, so in the current shape of hpx startstop backend I launch the backend for each parallel_for_() call and do not forward arc/argv - this decision was made in order not to introduce globals to opencv for now

12:43 <hkaiser> that will be very slow

12:47 <hkaiser> jakub_golinowski: here is what you can do to inject settings into the startup: https://gist.github.com/hkaiser/48c0b80e1d6dff9f14c378e7bf07b183

12:47 <hkaiser> this is taken from a project where things are done once durin gload time, your setup may vary

12:48 <jakub_golinowski> hkaiser, thanks - so for setting the config I just create the .hpx.ini file in the directory where the binary is

12:48 <hkaiser> please reload, I changed it slightly

12:48 <hkaiser> that may work as well

13:08 <M-ms> jakub_golinowski: I think it might be time to clean up the start/stop backend to a more efficient version

13:09 <M-ms> after you've shown that the hpx_main backend is pretty much as fast as the pthreads backend after tweaking parameters

13:09 <jakub_golinowski> M-ms, I also started having this thoughts, when I encountered this problem :D

13:09 <M-ms> did you have any luck with the perf scripts?

13:09 <M-ms> or do you have any numbers yet?

13:09 <M-ms> (any new ones...)

13:10 <jakub_golinowski> M-ms, so the update is as follows: I have downloaded all the dnn tests and run it for different backends

13:10 <M-ms> ok

13:10 <jakub_golinowski> for now hpx is doing really badly for those and for now I am not sure why

13:10 <jakub_golinowski> overnight I have run the mandelbrot with max_idle_loop_count=1

13:10 <M-ms> ok, was this again with default settings then?

13:11 <jakub_golinowski> but as you can infer from the discussion above the start-stop backend did not pickup the setting

13:11 <M-ms> yeah, well, we know already that one is slow so I think benchmarking it even further at this point is not necessary

13:12 <M-ms> if we can first show that the hpx_main backend is as fast as the others we can then start looking at the start/stop backend

13:12 <jakub_golinowski> https://drive.google.com/file/d/1WwrOr9Cy4cbYDWS-W-0_80chBjaRDrGt/view?usp=sharing

13:12 <jakub_golinowski> here is the plot

13:13 <M-ms> ok, it's closer but not quite there

13:14 <M-ms> also 4 threads?

13:14 <jakub_golinowski> yes, 4 threads

13:14 <M-ms> ok

13:14 <M-ms> and the dnn test was with what hpx backend and what settings?

13:14 <jakub_golinowski> M-ms, I tried different settings but no major improvements

13:15 <jakub_golinowski> let my try to use their summary.py to create a nice table

13:15 <M-ms> ok, thanks

13:15 <jakub_golinowski> M-ms, also you think I should make an issue about their errors in the report.py or do a PR right awat

13:16 <M-ms> yeah, either depending on how much you found out from the discussion yesterday

13:17 <M-ms> they'd probably appreciate a PR and can help if there's something they want done differently

13:17 <M-ms> that was my next question as well, do you have a gist/diff/something with what changes you needed in the end?

13:22 <jakub_golinowski> M-ms, wow they create a neat html: https://drive.google.com/open?id=1bvSv-uEXP2Xx8m4U3qDw20eBR4Kg_09s

13:28 <jakub_golinowski> M-ms, as for the python script that gave me errors: in the docu of the perf scripts they say that scripts are written for python 2 so I used pythnon 2.7 interpreter. Nevertheless it failed. After the investigation I found the reason which was not very well done patch to make one function forward compatible with python 3

13:29 <jakub_golinowski> However it seems strange to me that sb made this patch there since when I tried to use python 3 interpreter I get errors from some other places in the file

13:37 <jakub_golinowski> M-ms, here is a sketch of the changes: https://gist.github.com/Jakub-Golinowski/f14dd1b946ef3853adfca6a89f3ad29b

13:41 <M-ms> jakub_golinowski: thanks

13:42 <M-ms> so if your fix creates problems with python 3 I think it's better if you just open an issue

13:42 <M-ms> you can still include your suggested changes for python 2 but say that it doesn't work for python 3

13:42 <M-ms> that's a really useful table!

13:44 <M-ms> so there still seems to be a lot of variation, but in general the last hpx backend is not at all bad

13:44 <M-ms> again, we don't expect hpx to be able to beat the pthreads backend (at least in general) but here it's already faster in some cases

13:45 <M-ms> I've run the tests here and can post you my results as well once I have time (still today)

13:47 <jakub_golinowski> M-ms, I am not saying my fix introduces problems wiht python3. I am saying that when I try to run the script as is with the python3 interpreter it gives some errors anyway. Therefore a fix tha sb (not me) propose in function reformatTextValue() seems useless

13:47 <M-ms> aha, sorry

13:48 <jakub_golinowski> M-ms, also maybe I could get an access to some node? because running on my laptop becomes a bit of a problem cause I do not want to use it too much when tests run not to interefere with tests results?

13:48 <M-ms> so it has different errors with python 2 and 3, and yours fixes it for 2?

13:48 <jakub_golinowski> M-ms, exactly

13:48 <M-ms> in any case, at least open an issue and if they are happy with your suggested change you can make a PR out of it

13:49 <M-ms> jakub_golinowski: understandable... I can't promise anything

13:50 <M-ms> I think jbjnr got access to daint (at cscs) for his student

13:50 <jakub_golinowski> M-ms, actually now when I think about it I could try some ETH nodes - sudents are supposed to have acces to some cluster

13:51 <M-ms> true, you should have access to euler I think

13:51 <jakub_golinowski> M-ms, but do you now if the resources are guaranteed in euler when the job is started?

13:52 <M-ms> I've never used it but I assume they would use something like slurm which should give you unique access to a node

14:02 <nikunj> hkaiser, yt?

14:13 <hkaiser> nikunj: here

14:14 <nikunj> hkaiser, should I update the mac implementation now?

14:14 <nikunj> Or should I wait for for more linux tests

14:16 <hkaiser> nikunj: did you give up on the idea to improve diagnostics if startup goes wrong?

14:16 <nikunj> hkaiser, no, I'm thinking about that as well

14:17 <hkaiser> k

14:17 <nikunj> In the meantime we could update the code for mac

14:17 <hkaiser> sure, pls go ahead

14:17 <nikunj> hkaiser, ok

14:17 <hkaiser> nikunj: have you seen now that I passed you for GSoC?

14:18 <nikunj> yes

14:18 <hkaiser> k

14:47 <github> [hpx] K-ballo force-pushed logging from dde4b04 to a1538bc: https://git.io/vx6Yc

14:47 <github> hpx/logging a1538bc Agustin K-ballo Berge: pruning util/logging

15:45 <K-ballo> I'm doing a full build (for the first time in years), and hitting all kind of compilation errors in tests

15:45 <nikunj> K-ballo, is it related to my implementation?

15:46 <K-ballo> no, most look like they've been there forever

15:46 <K-ballo> there's even basic syntax errors, like missing semicolons

15:47 <K-ballo> here, for instance, https://github.com/STEllAR-GROUP/hpx/blob/master/hpx/lcos/local/spinlock_no_backoff.hpp#L77

15:48 <K-ballo> https://github.com/STEllAR-GROUP/hpx/commit/16e51e5f5c3b542b65bb9f45a099848a503077e9

15:49 <nikunj> And I used to think I was the only one missing out on semicolons :p

15:51 <K-ballo> some other files are missing includes

15:51 <K-ballo> and there's something going on with the lambda capture mechanism

15:52 <zao> On some "exotic" toolchain?

15:52 <K-ballo> msvc...

15:52 <zao> hehe

15:52 <K-ballo> /std:c++latest, but for some reason it looks like it is targetting C++11

15:53 <K-ballo> some project even ends up attempting to compile __asm__ blocks

16:06 <zao> You're not gonna trick me into building it, got other stuff to hack on :)

16:06 <K-ballo> it seems the HPX_CAPTURE_FORWARD behavior is bogus, mixing by-value captures with implicit by-value captures is ill-formed

16:10 <K-ballo> there's a bunch of std::uniform_int_distribution<std::uint8_t> too, apparently I introduced those, also ill formd

16:19 <hkaiser> K-ballo: got some errors?

16:19 <K-ballo> working on them

16:19 <hkaiser> K-ballo: perfect, thanks!

16:19 <K-ballo> we don't have msvc builders anymore?

16:19 <hkaiser> K-ballo: appveyor runs out of memory :/

16:20 <zao> I thought we had appveyor, but apparently ... ah

16:20 <hkaiser> K-ballo: ^^

16:23 <github> [hpx] K-ballo created compiler-fence-missing-semicolons (+1 new commit): https://git.io/fNmXX

16:23 <github> hpx/compiler-fence-missing-semicolons 9173773 Agustin K-ballo Berge: Add missing semicolons to uses of HPX_COMPILER_FENCE

16:27 <github> [hpx] K-ballo created uniform-int-distribution-byte (+1 new commit): https://git.io/fNmXH

16:27 <github> hpx/uniform-int-distribution-byte 56a6f9f Agustin K-ballo Berge: Replace undefined instantiations of uniform_int_distribution for std::uint8_t

16:48 <github> [hpx] K-ballo created fix-hpx-capture-default (+1 new commit): https://git.io/fNm1z

16:48 <github> hpx/fix-hpx-capture-default 1e13eba Agustin K-ballo Berge: Fix usage of HPX_CAPTURE together with default value capture [=]

16:49 <hkaiser> K-ballo: thanks!

18:08 <nikunj> zao: yt?

18:08 <zao> *ducks*

18:09 <nikunj> zao:could you please run my pr on Mac OS sometime when you're free. (https://github.com/STEllAR-GROUP/hpx/pull/3367)

18:09 <zao> nikunj: Nope. Left my mac at work.

18:10 <nikunj> oh, no worries.

21:21 <github> [hpx] K-ballo force-pushed fix-hpx-capture-default from 1e13eba to 3618f87: https://git.io/fNmdZ

21:21 <github> hpx/fix-hpx-capture-default 3618f87 Agustin K-ballo Berge: Fix usage of HPX_CAPTURE together with default value capture [=]

22:25 <K-ballo> CI builds take forever

22:56 <github> [hpx] K-ballo opened pull request #3380: Add missing semicolons to uses of HPX_COMPILER_FENCE (master...compiler-fence-missing-semicolons) https://git.io/fNmxk

22:57 <github> [hpx] K-ballo opened pull request #3381: Replace undefined instantiations of uniform_int_distribution (master...uniform-int-distribution-byte) https://git.io/fNmxt