############################
09:02:49	 De  Jean : will the video be uploaded on the website afterward ?
09:03:28	 De  France Boillod-Cerneux : yes, the vieo will be on Eurofusion youtube channel
09:03:38	 De  Jean : Thanks :)
09:04:26	 De  France Boillod-Cerneux : here is the link: https://www.youtube.com/channel/UCQNxXoQUPMOo_ETr09N3OtA
09:04:39	 De  France Boillod-Cerneux : you can find also all information about the webinar serie here:
09:04:51	 De  France Boillod-Cerneux : https://indico.euro-fusion.org/event/688/
############################
09:07:12	 De  sommariva : A small question, are there any compiler supporting standard 5.0 / 5.1 ?
09:09:43	 De  Michael Klemm : Not yet when it comes to full support of OpenMP 5.0/5.1, but most compiler vendors are adding features by priority, so you may see compilers supporting 5.1 features already while having incomplete 5.0 support.
The OpenMP ARB keeps a list of tools that we know about and for which the respective vendors provides us with information.  You can find that at https://www.openmp.org/resources/openmp-compilers-tools/
09:10:19	 De  Michael Klemm : You can find the cheat sheets here: https://www.openmp.org/specifications/
09:10:24	 De  sommariva : perfect, thank you very much for the info
09:10:55	 De  Michael Klemm : The spec book can be bought from Amazon at https://www.amazon.com/Application-Programming-Interface-Specification-Version/dp/B08NDZM8D4
09:11:27	 De  Michael Klemm : We also have a HTML version of the specification. The HTML version is linked from https://www.openmp.org/specifications/, too
09:22:28	 De  Michael Klemm : All of our examples are in C/C++, but the syntax transfers almost 1:1 to Fortran.  If you have a specific Fortran-related question, please let us know and we'll help.
############################
09:39:08	 De  Samuel Lazerson : Maybe I missed it but doesn’t the `i` variable need to be declared as private in the OMP for directive?
09:42:24	 De  Michael Klemm : If i is the loop iteration variable, then OpenMP automatically figures out that it needs to be private.  That only applies to the loop variable that OpenMP parallelizes over:

#pragma omp parallel for
for (i = 0; …) {…}

In that case, i is indeed private.  In this example:

#pragma omp parallel for
for(i = 0; …) {
  for (j = 0; …) {}
}

OpenMP will make i private, but j will remain as shared and thus needs a private(j) at the directive
09:53:45	 De  Michael Klemm : What Christian says about threadprivate also applies to Fortran COMMON blocks
############################
10:12:15	 De  Simppa Äkäslompolo : How well does unstructured parallelism work on GPUs? (Not very well in my experience, but maybe you can comment more.)
############################
10:13:54	 De  Tiago Ribeiro : slide4: does the implicit barrier play a role in these 2 examples?
10:15:54	 De  Christian Terboven : Well, there is also an implied barrier at the end of the parallel region.
10:17:11	 De  Tiago Ribeiro : thanks, but I guess this would make a difference if there was additional code inside that region, after the omp master, right?
10:18:10	 De  Christian Terboven : Yes, if there was other code after master / single, then it would make a difference.
############################
10:18:50	 De  Dimitrios Peponis (NCSRD/NKUA) : In a for loop, where a lot of race conditions happen due to concurrent writing in memory (more specific to array elements), I found that the critical 
clause is even slower than the serial execution of the code. In this case, is parallelization even possible?
10:20:59	 De  Christian Terboven : Dimitrios: the only real answer is "it depends". Sometimes it is impossible, but sometimes not. An example for the latter: if you use critical to ensure correctness, 
but if the likelyhood of conflicting accesses is rather low, then a transactional synchronization could help. You can add a "lock hint" to the OpenMP critical contruct.
10:21:20	 De  Samuel Lazerson : Can you put that in the chat so we can copy paste.
10:22:13	 De  victor cameo : https://github.com/cterboven/OpenMP-tutorial-EUROfusion
10:22:30	 De  Christian Terboven : e critical clause is even slower than the
10:22:44	 De  Christian Terboven : Sorry.
10:22:47	 De  Christian Terboven : Thank you, victor.
10:22:49	 De  Dimitrios Peponis (NCSRD/NKUA) : In fact, the code is rather large and involves a large number of different functions and source files. 
However, since you gave me the hint of lock hint and transactional synchronization I will try this. Could you please provide me with the documentation on these issues?
10:24:17	 De  Christian Terboven : #pragma omp critical hint(OMP_SYNC_HINT_SPECULATIVE)
10:24:33	 De  Dimitrios Peponis (NCSRD/NKUA) : Thank you!
10:24:50	 De  Michael Klemm : The sync hints are listed here: https://www.openmp.org/spec-html/5.1/openmpsu110.html#x143-1550002.19.12
10:25:07	 De  Michael Klemm : note, the "omp_lock_hint_" versions have been deprecated
10:25:23	 De  Michael Klemm : so, use the ones that start with "omp_sync_hint"
############################
10:26:22	 De  Alf Köhn-Seemann : Will OpenMP usuage (or rather support) on accelarators (GPUs) be addressed?
10:26:59	 De  Christian Terboven : Dimitrios: yes, we can do so.
############################
10:27:52	 De  Simppa Äkäslompolo : In april/may will we use m100 as the platform?
10:29:03	 De  Simppa Äkäslompolo : m100=marconi100=eurofusion supercomputer
10:28:23	 De  Michael Klemm : Not sure yet. It might be MI50.
10:29:05	 De  Michael Klemm : But you will be able to also use an IBM system with Nvidia GPUs or an Intel system using Intel oneAPI and a Gen GPU
############################
10:27:59	 De  France Boillod-Cerneux : https://indico.euro-fusion.org/event/688/
10:28:04	 De  France Boillod-Cerneux : here is the linki!
10:28:44	 De  France Boillod-Cerneux : https://www.youtube.com/channel/UCQNxXoQUPMOo_ETr09N3OtA
############################
10:25:24	 De  Simppa Äkäslompolo : thread safety of MPI on OpenMPI/MPI should be addressed?
Oral answer: YES
############################
10:26:35	 De  Dimitrios Peponis (NCSRD/NKUA) : Since in my case I have multiple write accesses to same array elements, I thought about using custom reduction clauses. Will you cover these in a future webinar?
Oral answer: YES
############################
10:24:14	 De  Samuel Lazerson : Will you cover hybrid OpenMP/MPI coding in a future Webinar?