- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Link Copied
- « Previous
-
- 1
- 2
- Next »
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Sergey,
Try cl with: #pragma omp parallel for schedule ( foobar )
Is cl quiet? If it is, then your report is no confirmation that schedule ( auto ) is supported by cl.
Jim Dempsey
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
>Sergey,
Thank you very much for your tests.
It's confirmed that, at least intel 13.1.0.149, schedule(auto) works well,
although I did not recieve the warning about #pragma...
BTW, I'm confusing related to which version of the openMP (3.0, 3.1, and/or 4.0) is used in the intel 13.1.1....
>TimP
Thank you very much for your valuable opinion.
>omp schedule(auto) is one of those facilities which were postponed
If so, unfortunately, should I wait a release of 14.0.0 or later...?
Or, schedule(runtime) and OMP_SCHEDULE seems to work fine for
what I want to do although many code in my appication are needed to be modified.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I'm expecting to see a more authoritative tabulation of which OpenMP features appeared when. It does appear that the 2013 update 2 and 3 compilers fairly well completed implementation of openmp 3.0. I'll try my own schedule(auto) case to see if it works with update 3 or some combination of compiling with one and run time shared libraries of the other, of which I've already seen an instance.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I ran a short benchmark including 3 exampes of schedule(auto) on a single architecture with all combinations of 2013 update 2,3, and beta compilers and the corresponding OpenMP libraries. It looks like schedule(auto) has worked well in such simple cases from update 2 on.
Object files made with update 3 didn't work with the other versions and had to be rebuilt.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
The case under discussion here seems to be one where an OpenMP 3.0 capability wasn't fully implemented until some 2013 update. collapse is another where important cases first worked in update 2.
max/min reduction hasn't yet worked for me, even though the parallel version of it is in OpenMP 3.1. An Intel presentation I saw recently claimed that it won't be standardized until OpenMP 4.0 is final.Of course, that's correct in reference to the #pragma omp simd reduction(max: which I don't think is available yet in any released or beta compiler.
In principle, #pragma omp parallel simd reduction could be used to simplify cases where it's currently best to write nested loops (outer parallel, inner vectorizable) with the best outer loop strategy differing according to the number of threads
#pragma omp simd safelen(32) is an OpenMP 4.0 feature which works better than the non-standard #pragma simd vectorlength, for the compiler versions which support it, which apparently go back to 2013 update 2. I'm wondering whether to post this update in my public version.
My colleagues have agreed with me that there is justification for the confusion mentioned above, and I'm hoping for some documentation to shed light on it. One would expect that interrogation of your OpenMP as to which version it supports would imply that all features of the reported version are present (at least to the extent that the application doesn't break when correct syntax is used). That hasn't been true of all past versions.
A full implementation of OpenMP 4.0 isn't advertised even for 14.0 compilers, but publicity for partial implementation to a greater extent than 2013 update 3 seems to be on the way.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Oh, #pragma omp simd safelen is ignored for some target settings such as -msse4.1 but observed for -mavx.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page
- « Previous
-
- 1
- 2
- Next »