[cfe-dev] [libc++] Working on the parallel STL algorithms

[Hal Finkel via cfe-dev]

Hi, Erik,

That's great!

Gor, Marshall, and I discussed this after some past committee meeting. 
We wanted to architect the implementation so that we could provide 
different underlying concurrency mechanisms; including:

    a. A self-contained thread-pool-based implementation using a 
work-stealing scheme.

    b. An implementation that wraps Grand Central Dispatch (for Mac and 
any other platforms providing libdispatch).

    c. An implementation that uses OpenMP.

The OpenMP-based provider is essential for interoperability in 
applications using OpenMP, GCD seemed most natural on Mac, and the 
self-contained implementation is available for anyone else. CUDA+UVM 
should be possible in the future as well. All of these mechanisms can 
export an interface for scheduling tasks, and we specifically discussed 
the interface having a "bulk enqueue" capability. This is essential for 
efficiently mapping onto several potential underlying technologies (e.g. 
OpenMP). You don't want to generate individual tasks for each item, and 
for underlying technologies that can do intelligent things, you want the 
underlying technology to make the coarsening decisions.

par_unseq should map, via pragmas, either Clang-specific and/or OpenMP, 
to enabling vectorizaton. As a longer-term exercise, we might need 
additional compiler extensions to make this work well for some 
algorithms (such as sorting), dealing with exceptions, etc.

  -Hal

On 05/15/2017 11:54 AM, Erik Pilkington wrote:
> Hello all,
>
> I’m an intern at Apple this summer, and my project is to work on the 
> C++17 parallel algorithms library. In a recent thread, Hal mentioned 
> [0] that he had spoken with Marshall about the design of the library. 
> I’m currently working on a prototype that is pretty similar to what 
> was committed into lib/Support in r302748. I’m wondering what you 
> folks talked about, and if any consensus wrt to design has been 
> reached. Even if not, I’d love to hear about what thoughts were 
> exchanged, or any reference material that could bring me up to speed!
>
> It might also be worth having a conversation about how to divide up 
> the work to be done here. Were any of you intending to start working 
> on the library?
>
> Thanks!
> Erik
>
> [0] http://lists.llvm.org/pipermail/llvm-dev/2017-May/112988.html
>

-- 
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20170516/cbc1e0ef/attachment.html>