[J3] Comments on 23-146

[Van Snyder]

On Mon, 2023-02-27 at 15:21 +0000, Jeff Hammond wrote:
One key motivation for ASYNC is to support asynchronous offload of DO CONCURRENT regions to GPUs and similar devices.  How do you propose to implement
 that using DO CONCURRENT, given that we have no way to tell compilers which type of parallelism to use, and fine-grain parallelism such as SIMD and SIMT, which are targeted by DO CONCURRENT by many compilers, are not capable of implementing the desired asynchrony.

I assume that an async construct or "task" (a la ada") would be
annotated somehow to indicate a desire to use a GPU. There is no reason
that each FORK of a PARALLEL construct could not be similarly
annotated. The difference between DO CONCURRENT and
PARALLEL...FORK...END PARALLEL is that each FORK does a different
thing. Some of them would be suitable for GPUs, some not.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.j3-fortran.org/pipermail/j3/attachments/20230227/912de751/attachment-0003.htm>