Ompt_callback_work with ompt_scope_end not always dispatched

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

Using nvc 23.1, host-only:

Although registered as ompt_set_always, the following code doesn’t dispatch ompt_callback_work at ws-loop-end:

#pragma omp parallel
{
    #pragma omp for
    for (int i=1; i<5; i++)
        foo(i);
}

Expected events are:

encountering: parallel_begin
worker: implicit_task( begin )
worker: work( begin )
worker: work( end ) <- expected but missing
# sync for
worker: sync_region( begin )
worker: sync_region_wait( begin )
worker: sync_region_wait( end )
worker: sync_region( end )
# sync parallel
worker: sync_region( begin )
worker: sync_region_wait( begin )
worker: sync_region_wait( end )
worker: sync_region( end )
worker: implicit_task( end )
encountering: parallel_end

Note that the combined construct

#pragma omp parallel for
for (int i=1; i<5; i++)
    foo(i);

dispatches all events as expected, i.e.

encountering: parallel_begin
worker: implicit_task( begin )
worker: work( begin )
worker: work( end )
worker: sync_region( begin )
worker: sync_region_wait( begin )
worker: sync_region_wait( end )
worker: sync_region( end )
worker: implicit_task( end )
encountering: parallel_end

It would be nice if this could be fixed in the next release.

Issue also mentioned here: NVHPC 22.11/23.1 -- OMPT methods can cause SegFault when offloading - #6 by jan.andre.reuter

Thanks,
Christian

2

Thanks Christian. Do you have the complete example with the callbacks included? We can try to pull something together, but if you have it already, that would be great.

-Mat

3

ompt-printf-0.tar.gz (122.0 KB)
reproducer2.c (245 Bytes)
reproducer1.c (289 Bytes)

Hi Mat,

please use reproducer1.c and reproducer2.c with the attached ompt-printf-0.tar.gz like this:

tar xf ompt-printf-0.tar.gz
cd ompt-printf-0
./configure CC=nvc --prefix=`pwd`/_install
make install
# See make install's compile and link instructions and build reproducer1 and reproducer2. DOn't forget to add -mp=ompt to the link line.

The output is more verbose than what I posted originally (use low thread count), but it should be straight forward to decode.

Thanks
Christian