On 1/20/20 12:07 PM, Jakub Jelinek wrote:
I'd say easiest would be to do that in the gcn specific mkoffload. But
there needs to be a way for the user to specify that he wants only a
particular variant and not all of them (perhaps look for -march= in
the offload options?)?
I think that relates to the general issues with telling libgomp what
offload-target has been used.
(Might be solved differently, but it is at least related. Thinking of
complications would be: main program and a linked library uses different
offload targets [e.g. nvptx vs. host or gcn vs. host+nvptx] or the
library is not linked but dlopen'ed. — Likewise for the issue of this
thread: library and host program might use different gfx… for the march.)
* * *
Regarding the -foffload=<target(s)>: Assume a GCC installation which has
plugins for a certain device. If one now uses a simple hello-world
program with OpenMP target or OpenACC then:
With "-fopenacc -foffload=disable" – and a device available – it fails
at run time with "libgomp: target function wasn't mapped". (Work around:
set env var ACC_DEVICE_TYPE to "host".)
With "-fopenmp -foffload=disable" or "-fopenacc -foffload=disable" – and
no device available, it fails with: libgomp: while loading
libgomp-plugin-nvptx.so.1: libcuda.so.1: cannot open shared object file:
No such file or directory
And – at least with HSA/GCN: If there is a permission issue for /dev/…,
it will fail when the devices are enumerated (I mean:
gomp_init_targets_once → *get_num_devices) – even if -foffload=disable
is used.
See also: PR 81886 and, related, PR 67300.
Cheers,
Tobias
