[cfe-dev] Comparison of 2 schemes to implement OpenMP 5.0 declare mapper codegen

[James Beyer via cfe-dev]

Recursive data structures are important if you consider linked lists important.

Supporting these is challenging but not impossible, I would expect that if someone manages to implement a cost effective way to support linked lists we would add support to OpenMP with ease.

From: Finkel, Hal J. <hfinkel at anl.gov>
Sent: Friday, June 28, 2019 10:46 PM
To: Alexey Bataev <Alexey.Bataev at ibm.com>; Li, Lingda <lli at bnl.gov>
Cc: Alexandre Eichenberger <alexe at us.ibm.com>; Chapman, Barbara (Contact) <barbara.chapman at stonybrook.edu>; Kevin K O'Brien <caomhin at us.ibm.com>; Carlo Bertolli <cbertol at us.ibm.com>; Deepak Eachempati <deachempat at cray.com>; Denny, Joel E. <dennyje at ornl.gov>; David Oehmke <doehmke at cray.com>; Ettore Tiotto <etiotto at ca.ibm.com>; fraggamuffin at gmail.com; Rokos, Georgios <georgios.rokos at intel.com>; Gheorghe-Teod Bercea <Gheorghe-Teod.Bercea at ibm.com>; gregory.rodgers at amd.com; Sharif, Hashim <hsharif3 at illinois.edu>; Cownie, James H <james.h.cownie at intel.com>; Sjodin, Jan <Jan.Sjodin at amd.com>; James Beyer <jbeyer at nvidia.com>; Doerfert, Johannes <jdoerfert at anl.gov>; Jones, Jeff C <jeff.c.jones at intel.com>; josem at udel.edu; Robichaux, Joseph <joseph.robichaux at intel.com>; Jeff Heath <jrheath at ca.ibm.com>; khaldi.dounia at gmail.com; Kelvin Li <kli at ca.ibm.com>; Bobrovsky, Konstantin S <konstantin.s.bobrovsky at intel.com>; Kotsifakou, Maria <kotsifa2 at illinois.edu>; Li, Lingda (Contact) <lildmh at gmail.com>; Lopez, Matthew Graham <lopezmg at ornl.gov>; lopezmg at ornl.org; Menard, Lorri <lorri.menard at intel.com>; Martin Kong <martin.richard.kong at gmail.com>; Sarah McNamara <mcnamara at ca.ibm.com>; Rice, Michael P <michael.p.rice at intel.com>; Matt Martineau <m.martineau at bristol.ac.uk>; oscar at ornl.gov; Jeeva Paudel <pjeeva01 at ca.ibm.com>; Rao, Premanand M <premanand.m.rao at intel.com>; Krishnaiyer, Rakesh <rakesh.krishnaiyer at intel.com>; Narayanaswamy, Ravi <ravi.narayanaswamy at intel.com>; Monteleone, Robert <robert.monteleone at intel.com>; Lieberman, Ron <Ron.Lieberman at amd.com>; Samuel Antao <Samuel.Antao at ibm.com>; Jeffrey Sandoval <sandoval at cray.com>; Sunita Chandrasekaran <schandra at udel.edu>; sergey.y.ostanevich at gmail.com; Sergio Pino Gallardo <sergiop at udel.edu>; Dmitriev, Serguei N <serguei.n.dmitriev at intel.com>; Chan, SiuChi <siuchi.chan at amd.com>; Sunil Shrestha <sshrestha at cray.com>; Wilmarth, Terry L <terry.l.wilmarth at intel.com>; Tianyi Zhang <tzhan18 at lsu.edu>; vadve at illinois.edu; Wang Chen <wdchen at ca.ibm.com>; Wael Yehia <wyehia at ca.ibm.com>; Tian, Xinmin <xinmin.tian at intel.com>; cfe-dev at lists.llvm.org; openmp-dev at lists.llvm.org
Subject: Re: Comparison of 2 schemes to implement OpenMP 5.0 declare mapper codegen


Hi, Alexey, Lingda,

I haven't been following this closely, so a few questions/comments:

 1. Recursive mappers are not supported in OpenMP 5, but do we expect that to change in the future?

 2. Our experience so far suggests that the most important optimization in this space is to limit the number of distinct host-to-device transfers (or data copies) on systems where data needs to be copied. In these schemes, where does that coalescing occur?

 3. So long as the mappers aren't recursive, I agree with Alexey that the total number of to-be-mapped components should be efficient to calculate. The counting function should simplify to a trivial expression in nearly all cases. The only case where it might not is where the type contains an array section with dynamic bounds, and the element type also has a mapper with an array section with dynamic bounds. In this case (similar to the unsupported recursive cases, which as an aside, we should probably support it as an extension) we could need to walk the data structure twice to precalculate the number of total components to map. However, this case is certainly detectable by static analysis of the declared mappers, and so I think that we can get the best of both worlds: we could use Alexey's proposed scheme except in cases where we truly need to walk the data-structure twice, in which case we could use Lingda's combined walk/push_back scheme. Is there any reason why that wouldn't work?

Thanks again,

Hal
On 6/28/19 9:00 AM, Alexey Bataev wrote:

Hi Lingda, thanks for your comments.
We can allocate the buffer either by allocating it on the stack or calling OpenMP allocate function.
With this solution, we allocate memory only once (no need to resize buffer after push_backs) and we do not need to call the runtime function to put map data to the buffer, compiler generated code can do it.
But anyway, I agree, it would be good to hear some other opinions.
--------------
Best regards,
Alexey Bataev



...

--

Hal Finkel

Lead, Compiler Technology and Programming Languages

Leadership Computing Facility

Argonne National Laboratory

-----------------------------------------------------------------------------------
This email message is for the sole use of the intended recipient(s) and may contain
confidential information.  Any unauthorized review, use, disclosure or distribution
is prohibited.  If you are not the intended recipient, please contact the sender by
reply email and destroy all copies of the original message.
-----------------------------------------------------------------------------------
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20190629/4dedee24/attachment.html>