Hi Julian! On 2019-10-21T16:14:11+0200, I wrote: > On 2019-10-03T09:35:04-0700, Julian Brown <jul...@codesourcery.com> wrote: >> This patch has been broken out of the patch supporting OpenACC 2.6 manual >> deep copy last posted here: >> >> https://gcc.gnu.org/ml/gcc-patches/2018-12/msg01084.html > > Thanks.
I meanwhile re-discovered that an earlier submission, <http://mid.mail-archive.com/cover.1543438190.git.julian@codesourcery.com>, had included some documentation/rationale for: > I haven't understood all the changes related to replacing > 'dynamic_refcount' with 'virtual_refcount', getting rid of > 'data_environ', the 'lookup_dev' rework, but I trust you got that right. > In particular, these seem to remove special-case OpenACC code in favor of > generic OMP code, which is good. ... these changes. Please in the future remember to refer to such existing documentation/rationale, or again include in any re-submissions, thanks. >> Tested with offloading to NVPTX, with good results I noticed that when testing with '-foffload=x86_64-intelmicemul-linux-gnu', the x86_64-pc-linux-gnu '-m32' multilib (but not default '-m64', huh) then reproducibly regresses: PASS: libgomp.c/target-link-1.c (test for excess errors) [-PASS:-]{+FAIL:+} libgomp.c/target-link-1.c execution test ..., with an un-helpful message: "offload error: process on the device 0 unexpectedly exited with code 0". So non-OpenACC code paths seem to be negatively affected in some way? Hopefully that'll go away when backing out the 'VREFCOUNT_LINK_KEY' etc. changes, as discussed elsewhere. (I can easily test patches for you, no need for you to set up Intel MIC (emulated) offloading testing.) Grüße Thomas
signature.asc
Description: PGP signature
