Using -foffload=[...] to cycle through accelerators (was: libgomp nvptx plugin: rework initialisation and support the proposed load/unload hooks)

[Thomas Schwinge <thomas@codesourcery.com>]

[-- Attachment #1: Type: text/plain, Size: 1736 bytes --]

Hi!

On Tue, 14 Apr 2015 15:15:02 +0100, Julian Brown <julian@codesourcery.com> wrote:
> The other problem is that it appears that the ACC_DEVICE_TYPE
> environment variable is not getting set properly on the target for (any
> of) the OpenACC tests: this means a lot of the time the "wrong" plugin
> is being tested, and means that the above tests (and several others)
> still fail. That will apparently need some more engineering (on our
> part).

This should be working fine for "local" testing (that is, build-tree
testing, without a remote board).  Setting/communicating environment
variables to remote boards (which we use a lot for our internal testing)
indeed does not work in DejaGnu.  This has been reported several times in
the last years, by different people, but nobody ever came up with a
solution that was sufficiently generic so that was acceptable for
inclusion.  (Need a way to specify whether environment variables should
be set for the host and/or target system, and so on.)


For the problem at hand, I once had a different suggestion, which I'm
paraphrasing here: the existing code should be changed such that when
-foffload=nvptx,intelmic,... is specified (also considering a comple-time
default value based on --enable-offload-targets=[...]), the first
offloading target (nvptx in my example) is the one that is used by
default (acc_device_default) with OpenACC, and for OpenMP as device ID 0,
and so on.  That way, we could compile the libgomp test cases with
-foffload=$offload_target_openacc (see
libgomp/testsuite/libgomp.oacc-c/c.exp for context), and wouldn't have to
care about the ACC_DEVICE_TYPE environment variable anymore.

Does that make sense?


Grüße,
 Thomas

[-- Attachment #2: Type: application/pgp-signature, Size: 472 bytes --]