by

Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
  • Aug 05 15:38
    DzmitryFil edited #781
  • Aug 05 11:08
    DzmitryFil opened #781
  • Aug 04 10:11
    nikomatsakis synchronize #746
  • Aug 03 19:55
    DzmitryFil closed #780
  • Aug 01 20:18
    DzmitryFil edited #780
  • Aug 01 19:44
    DzmitryFil opened #780
  • Jul 29 22:03
    couchand closed #779
  • Jul 29 19:20
    cuviper labeled #779
  • Jul 29 18:05
    couchand opened #779
  • Jul 24 03:07
    cuviper closed #776
  • Jul 21 19:46
    bors[bot] closed #771
  • Jul 21 19:46
    bors[bot] closed #770
  • Jul 21 19:46
    bors[bot] closed #770
  • Jul 21 19:46

    bors[bot] on master

    par_iter for Range<char> par_iter for RangeInclusive<cha… Allow char iter tests to run on… and 4 more (compare)

  • Jul 21 19:39

    bors[bot] on staging.tmp

    (compare)

  • Jul 21 19:39

    bors[bot] on staging

    par_iter for Range<char> par_iter for RangeInclusive<cha… Allow char iter tests to run on… and 4 more (compare)

  • Jul 21 19:39

    bors[bot] on staging.tmp

    par_iter for Range<char> par_iter for RangeInclusive<cha… Allow char iter tests to run on… and 4 more (compare)

  • Jul 21 19:39

    bors[bot] on staging.tmp

    [ci skip][skip ci][skip netlify] (compare)

  • Jul 18 14:05
    v1gnesh opened #778
  • Jul 16 19:48
    bors[bot] closed #777
Josh Stone
@cuviper
if you don't care about order, maybe some kind of concurrent vector would be faster, which you could push in a for_each, but I don't have specific recommendations
flat_map is kind of awkwardly handled in rayon, since we split that in parallel too -- it may over-split your data and make it move data a lot to collect it back
sometimes you may do better with a custom fold/reduce
Zhengyi Yang
@zhengyi-yang
@cuviper Got it, thanks Josh
Niko Matsakis
@nikomatsakis
Man I'm so sad I apparently never pushed to https://github.com/nikomatsakis/rayon-plot
I wonder if some old computer of mine might have a checkout
I was looking @cuviper just a bit into the sort performance but nothing jumped out at me yet
Christopher N. Hesse
@raymanfx
Hi everyone. I'm looking for feedback similar to the case of rayon-rs/rayon#699, but didn't want to necro that old "issue". Basically, I'd like to parallelize the following (outer) loop: https://github.com/raymanfx/ffimage/blob/next/src/packed/convert.rs#L18. Now I'm not sure how to do that. I implemented iterators that hand out mutable pixel references, so I could iterate through the target image buffer. But then I don't have the pixel coordinates anymore which I need to transform the source pixel. I tried zipping the mutable target and the immutable source pixel iterators, but did not succeed that way. I'd appreciate if someone could give me a hint here.
  • adding to the above: I'd like to perform the conversion in memory, so don't want to collect the converted pixels in a Vec and then overwrite the image pixels with the ones from that vec
Niko Matsakis
@nikomatsakis
@cuviper so do you have any sense for how we ought to proceed with "the branch"?
I'm feeling a bit stuck :/
I dont' feel great about the perf results, it'd probably be good to drill a bit more into which ones we care about
I guess if you're interested, I might be game to schedule some time to screen share and do a bit of joint investigation :)
not sure if that's something that "pairs well"
also would Tuesdays at 15:30 Eastern time (12:30 Pacific) work for you as a sync slot?
Josh Stone
@cuviper
@nikomatsakis sorry I missed this. Tuesdays at 12:30 Pacific will be fine for me
for the branch, I don't know -- I'm almost inclined to just SHIP IT :rocket:
but we should probably make sure that it at least still improves the idle case it's targeting
(make sure the redesigns haven't lost that property)
Niko Matsakis
@nikomatsakis
@cuviper ok let me add something to the calendar. Regarding the branch, I kind of feel the same way.
Like, let's land it, maybe do an alpha release and encourage people to test
but yeah let's verify that we're solving the problem we set out to solve first I guess
I definitely feel like it'd be awesome if we could draw some eyes, maybe some poeple can come up in and optimize things
living on a PR is no way to live
Luca Barbato
@lu-zero
I can test it today
btw is there a way to move from a parallel iterator to a serial iterator without having a collect() in between?
Josh Stone
@cuviper
@lu-zero there's no builtin way to go parallel to serial. A channel might work, if you don't care about order.
Luca Barbato
@lu-zero
I care about order thus why I wanted that kind of adaptor ^^
meanwhile I have some data on ARM, our usual test encode is just slightly faster there
Josh Stone
@cuviper
well the issue for serial conversion is #210
faster is good :)
Luca Barbato
@lu-zero
:)
it seems so
more data tomorrow
and I will try on power as well
Luca Barbato
@lu-zero
On another machine I see no difference between the two for a similar test
Luca Barbato
@lu-zero
Setting delta rayon
Speed 10 Tiles 4 0.42%
Speed 10 Tiles 8 0.85%
Speed 10 Tiles 16 6.93%
Speed 8 Tiles 4 0.09%
Speed 8 Tiles 8 0.71%
Speed 8 Tiles 16 4.37%
Speed 6 Tiles 4 0.26%
Speed 6 Tiles 8 0.68%
Speed 6 Tiles 16 4.46%
This on x86
Luca Barbato
@lu-zero
Setting nightly stable
Speed 10 Tiles 4 0.42% -0.17%
Speed 10 Tiles 8 0.85% 1.75%
Speed 10 Tiles 16 6.93% 6.27%
Speed 8 Tiles 4 0.09% -0.36%
Speed 8 Tiles 8 0.71% 0.73%
Speed 8 Tiles 16 4.37% 1.71%
Speed 6 Tiles 4 0.26% 0.27%
Speed 6 Tiles 8 0.68% 0.95%
Speed 6 Tiles 16 4.46% 2.05%
Now both with current nightly rust and stable
Luca Barbato
@lu-zero
Setting    delta nightly    delta stable
Speed 10 Tiles 4    12.68%    12.88%
Speed 10 Tiles 8    14.85%    15.30%
Speed 10 Tiles 16    16.37%    16.92%
Speed 8 Tiles 4    8.18%    8.57%
Speed 8 Tiles 8    10.93%    11.19%
Speed 8 Tiles 16    12.47%    13.16%
Speed 6 Tiles 4    7.83%    8.21%
Speed 6 Tiles 8    10.77%    11.31%
Speed 6 Tiles 16    12.36%    12.76%
(I let somebody else to format it)
on arm the impact is crazy
Josh Stone
@cuviper
are these percentages faster or slower?
Luca Barbato
@lu-zero
faster
Josh Stone
@cuviper
whew
I was worried
Luca Barbato
@lu-zero
I wonder if there is something wrong with those or with the other arm box
I'd doublecheck them
Josh Stone
@cuviper
what are the sizes (#cpus) of the machines you're testing?