This is done to provide the functionality when the compiler doesn't support inline asm. I do not know how testing is done with the atomics in opal/asm/base atomics so its possible some of them are incorrect.
-Nathan On Fri, Apr 01, 2016 at 02:39:39PM +0530, Sreenidhi Bharathkar Ramesh wrote: > (Sorry, typo in previous email. Please read that as) > I have another question: why is any given operation implemented twice for > the same architecture? > Ref: opal/asm/base/AMD64.asm and opal/include/opal/sys/amd64/atomic.h, > say function opal_atomic_cmpset_32() > Thanks! > On Fri, Apr 1, 2016 at 2:26 PM, Sreenidhi Bharathkar Ramesh > <sreenidhi-bharathkar.ram...@broadcom.com> wrote: > > Hi Paul, > Thanks for the response, I understand. > I have another question: why is any given operation implemented twice > for the same architecture? > Ref: opal/asm/base/ARM.asm and opal/include/opal/sys/amd64/atomic.h > Please let me know. > Thanks, > - Sreenidhi. > On Thu, Mar 31, 2016 at 2:28 AM, Paul Hargrove <phhargr...@lbl.gov> > wrote: > > Sreenihdi, > Maintenance of the "other" files is, at least, required for older > compilers (including several non-GNU ones) which have support for GCC > inline ASM constructs but do not include the __sync functions. > Additionally, the __sync functions always include a full memory fence > while there are places where Open MPI only needs Acquire or Release. > In that sense the use of the GCC builtin atomics is not always an > "optimization". > -Paul > On Wed, Mar 30, 2016 at 4:19 AM, Sreenidhi Bharathkar Ramesh > <sreenidhi-bharathkar.ram...@broadcom.com> wrote: > > Hello, > I did few more experiments. Built-in atomics are disabled by > default and can be enabled by "--enable-builtin-atomics". With this > option, a different atomic file is include (I mean, different from > the other three mentioned in the earlier email). > opal/include/opal/sys/sync_builtin/atomic.h > By enabling "--enable-builtin-atomics", the OpenMPI library will > not only get the latest optimizations from GCC __sync functions, but > also get all additional functions, such as <__sync_or_and_fetch> > <__sync_and_and_fetch> etc. Also, this way, the arch/*.asm files > need not be maintained. > Seems there is more of an advantage to have > "--enable-builtin-atomics". Am I missing something ? Please let me > know your thoughts. > Thanks, > - Sreenidhi. > On Wed, Mar 30, 2016 at 10:33 AM, Sreenidhi Bharathkar Ramesh > <sreenidhi-bharathkar.ram...@broadcom.com> wrote: > > Hello, > With respect to atomic operations in assembly, I can see three > source files, for the same AMD64 architecture. > 1. opal/asm/base/AMD64.asm > 2. opal/include/opal/sys/amd64/atomic.h > 3. opal/include/opal/sys/atomic_impl.h > What is the need for three files ? How are they inter-related ? > Please let me know. > Thanks, > - Sreenidhi. > > _______________________________________________ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2016/03/18733.php > > -- > Paul H. Hargrove phhargr...@lbl.gov > Computer Languages & Systems Software (CLaSS) Group > Computer Science Department Tel: +1-510-495-2352 > Lawrence Berkeley National Laboratory Fax: +1-510-486-6900 > _______________________________________________ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2016/03/18734.php > _______________________________________________ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2016/04/18737.php
pgpBbz7xnn5Cm.pgp
Description: PGP signature
