Le 2016-03-13 16:22, Thakur, Rajeev a écrit :
In the 19 years since the MPI-2 standard was released, you may be the first one requesting this functionality :-). I am curious what are you using it for? Here is the reason....
In 2012, when I read your book about MPI-2, I wrote all our I/O functions to use collective MPI I/O calls with "native" as datarep. I knew it would not be portable "as is" but the main usage is 64bits native format... then I told myself "It will be ok when I will want to read/write from 32 bits architectures or different endian style, since there is that user defined datarep possibility, as illustrated in chapet 9.1 of the book...", but unfortunately for me, I didn't _tested_ the functionality in 2012... :/ So we moved all our file formats (meshes, results, etc) to be aligned with the output of "native" datarep into a 64bits standard linux x86 architecture... That means we have produced, for example, 1200 validations tests with input/output files into that format... That is, of course, excluding all user data since 4 years.. No I am in the phase of porting the code to fully support 32bits architecture (we have an 32bit ubuntu running here, and working on windows port). So, I did test for the first time the datarep functionality to be able to convert 64 bits "long int" from/to 32 bits "long int" for writing/reading in respect to existing file format... It was all an easy task in my mind with the user datarep functionality just as it is in the standard.. But it appears I am the only one in the world (wow...) to request a way to have his files portable with MPI... without performance losses for the "main" usage which is 64bits... Maybe there is another simple way to do this? Anyway, I am now expecting that you will tell me it will never be implemented _just for me_... :/ Maybe I can contribute? :) Thanks, Eric
Rajeev
On Mar 13, 2016, at 1:51 PM, Eric Chamberland <[email protected]> wrote:
Le 2016-03-10 18:04, Thakur, Rajeev a écrit :
Support for user-defined data representations has not been added to MPICH yet. There has not been much demand for it. Thanks for your answer...
Is it a "big" task?
Should it be scheduled so it may appear in a forthcoming version?
Thanks,
Eric
_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss
discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss
_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss