Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
  • 05:13
    wave-b commented #152
  • 05:02
    freemo closed #153
  • 05:02
    freemo commented #153
  • 03:27

    freemo on master

    Attempt at a middleman CI for d… Updating middleman Added back live reload. and 14 more (compare)

  • 03:20

    freemo on v2.x

    (compare)

  • 03:15

    freemo on v2.x

    (compare)

  • 03:15

    freemo on master

    Fix: Issue #153 JVM Crash - Mul… Fix: Issue #153 JVM Crash - Mul… Refactor: Refactored ArrayBuffe… and 4 more (compare)

  • 03:09

    freemo on release

    release 1.2.0 (compare)

  • 03:03

    freemo on v1.4.x

    (compare)

  • 03:03

    freemo on master

    docs(javadocs): fixed all JavaD… Merge branch 'master' into 'mas… (compare)

  • 02:47
    freemo closed #152
  • 02:47
    freemo commented #152
  • 02:37
    Travis Syncleus/aparapi (master) errored (385)
  • 02:32

    freemo on release

    (compare)

  • 02:28
    Travis Syncleus/aparapi@f4869bd (release) errored (384)
  • 02:27

    freemo on win-fix

    (compare)

  • 02:27

    freemo on v2.x

    (compare)

  • 02:27

    freemo on master

    Updated the Maven POM to use th… Merge branch 'master' into 'mas… Test: Add test for #153 - IntAr… and 2 more (compare)

  • 02:27

    freemo on release

    Release v1.5.0 (compare)

  • Aug 22 12:57
    CoreRasurae commented #152
CoreRasurae
@CoreRasurae
I haven't registered yet
Jeffrey Phillips Freeman
@freemo
@CoreRasurae I dont think so, you need an account on our gitlab instance is all but not permission I dont think: https://git.qoto.org
@CoreRasurae the reason for th emove to gitlab is mostly for the CI since i can run GPU enabled runners for CI. Something that last i checked travis doesnt offer. Plus I like the gitlab CI system overall and many other features.
CoreRasurae
@CoreRasurae
sure
Jeffrey Phillips Freeman
@freemo
@CoreRasurae but as i said it will all mirror back to github anyway
CoreRasurae
@CoreRasurae
cool
no problem
Jeffrey Phillips Freeman
@freemo
let me know if it gives you any issues, registration should be open to all
CoreRasurae
@CoreRasurae
is aparapi-Native on gitlab?
too
Jeffrey Phillips Freeman
@freemo
Should be let me double check
yup, it and all the other projects are int he Aparapi group: https://git.qoto.org/aparapi?nav_source=navbar
CoreRasurae
@CoreRasurae
ok
Jeffrey Phillips Freeman
@freemo
CoreRasurae
@CoreRasurae
@freemo Just submitted MergeRequest for CI test
@freemo tomorrow will add the fix in aparapi-native
Jeffrey Phillips Freeman
@freemo
@CoreRasurae ohh cool! I might check it out tonight, but half falling asleep so may be tomorrow.
CoreRasurae
@CoreRasurae
@freemo I have Aparapi and Aparapi-Native Merge Requests ready.
@freemo Only Aparapi-Native has code modification to fix the issue and a small refactoring with no functionality change
Jeffrey Phillips Freeman
@freemo
@CoreRasurae Awesome ill review shortly.
CoreRasurae
@CoreRasurae
@freemo I also gave a look at #152 and I believe we can close it
@freemo It is simply something that is really unsupported in Aparapi
@freemo To have multi-dimensional arrays where inner arrays dimensions vary with the index of the outer dimension
Jeffrey Phillips Freeman
@freemo
@CoreRasurae would it make more sense to just tag it as a feature request instead and let it sit int he backlog incase we decide to add the feature in the future?
CoreRasurae
@CoreRasurae
@freemo Well if you want to, but I think it something that doesn't make sense, it is simply inefficient in terms of memory accesses, to which GPUs are quite sensitive, and multi-dimensional arrays are already inefficient because OpenCL 1.x can only deal with 1D arrays, so that Aparapi needs to convert them forth and back at each kernel execution.
@freemo It is something unusual to have an array like: int x[][][] = new int[2]; x[0] = new int[100][200]; x[1]=new int[200][100];
@freemo It would also impact OpenCL kernels because the kernel would have to keep this variable dimension mapping by index, which would make the kernel really slow
Jeffrey Phillips Freeman
@freemo
@CoreRasurae ill review it in more detail in a few minutes and we can chat then. Just need to eat something first
CoreRasurae
@CoreRasurae
ok
I'm also working on other project now
but feel free to comment, I'll be around
Jeffrey Phillips Freeman
@freemo
@CoreRasurae alright
CoreRasurae
@CoreRasurae
@freemo Did you had a chance to take a look?
Jeffrey Phillips Freeman
@freemo
@CoreRasurae looking as we speak
@CoreRasurae so far everything looks good, ill approve. I also agree tha tthe ticket doesnt make sense as a feature. Ill labe it as a feature we wont do
@CoreRasurae ill run the tests manually for now. You tested on linux only i presume?
CoreRasurae
@CoreRasurae
@freemo Yes, only on Linux, on my laptop, which is what I have here. BTW I can also test on ARM hard float 32-bits, with Linux and an AMD graphics card
CoreRasurae
@CoreRasurae
@freemo Aparapi can also work on armhf 32-bits and 64-bits, since this arch is becoming popular, I think it would be nice to add support for it. I tested aparapi-native and it compiles cleanly, it is just a matter of adapting aparpi-jni to load the libraries accordingly.
Jeffrey Phillips Freeman
@freemo
@CoreRasurae I will setup an arm machine int he CI sometime this week
just trying to manually test for this one merge
Should have it merged soon just need to get my windows box back up
Doing that now and will be merged later today if it passes tests
CoreRasurae
@CoreRasurae
Ok, great news, both of them
@freemo cool
jchan007
@jchan007
@freemo Looking forward to the merge. Hoping it might resolve the issue I ran across.
grfrost
@grfrost
@CoreRasurae and @freemo I agree with Jeffrey here. We tried to allow this in the original Aparapi, but dropped it. It makes coding easier but is very inneficient. Ideally you can replace this with a single array (100x200)+(200x100) in length and manage the indices yourself. This would only prove viable on a HSA compatable device with fully coherent memory. The other advantage of a single block of memory is that it is less likely to be moved by the garbage collector. The bigger the object the more likely it will be allocated (or quickly moved to) perm-gen.
CoreRasurae
@CoreRasurae
@jchan007 The code has already been merged. You can find it in: https://git.qoto.org/aparapi/aparapi-native
@grfrost I also agree with what you say
Jeffrey Phillips Freeman
@freemo
@jchan007 it was merged let me know if it helped
Jeffrey Phillips Freeman
@freemo
@grfrost when you have some time send me a message. I'd like to catch up.