WIP for openacc (GPU) for MulticolvarTemplate

[Iximiel]

I am opening this PR to have a clearer diff and to make a few checks

In particular I need to understand why some tests do not pass anymore the forces part. I did not touch that part yet... (could be that skipping of the pbc passage 🤔)
So I need a clearer diff.

At the time of writing this the /targetGPU/rt-matrixsum that is adjmat/rt-matrixsum passes the "value" test but not the "force test"

If you try to compile/modify this remember that the dependency part of the build tree is somehow broken with nvc++ (at least with mine) and as now I am not trying to fix it. So to compile after a modification you have to manually delete the interested files (like this rm src/colvar/Distance.o ; make -j12)

[Iximiel]

@GiovanniBussi @gtribello
As it is now it seems to work, IF you do not look at the virial

Now if you run the new targetGPU/rt-basic-print/, the one with the USEGPU flags, it passes 😄, I think the forward loop is ok

The next thing to do (after finding a solution for the virial) will be documenting what I did, since working with openacc feels more or less like playing with an escherian jenga tower

[gtribello]

@Iximiel, the stuff for the secondary structure using the new ParallelTaskManager is in the branch called gpuwork/secondarystructure

[gtribello]

Hi @Iximiel

Further to my message the code for the new secondary structure stuff is in the gpuwork/secondarystructure-2

This version of the branch was built off final-backpropegation

[Iximiel]

I managed to make also the virial work 😄

as now is not mergeable, becasue we are spending more time in printf-debugging that other. But I wanted to show that the code was working.

I would like to do a small change while I a prepare this to be merged:
@gtribello do you mind if I change the name of the variable nderivatives_per_task into nderivatives_per_component in the ParallelTaskManager? It is more clear, and it was the casue of the confusion in the error of the virial.

I have already did something similar in ColvarOutput.h (where I also unrolled and removed the memory copies in setBoxDerivativesNoPbc, becasue I tought that there where a problem there...), now it it slighly clear, I think with the names of the variables.

[gtribello]

Hi @Iximiel

I made that change of variable name in the final-backpropegation branch that I pushed yesterday. If you have other changes that you think should be incorporated from the work you have done here then please PR them onto that branch.

Thanks
Gareth

[Iximiel]

The changes in configure.ac in 37388af make working with nvhpc compiler much much easier

[Iximiel]

rebased on final-backpropagation, need to adapt it to the new changes

[Iximiel]

@gtribello it works (at least on my PC), on the branch more aligned with the final-backpropagation

I have included the change I proposed here

I think I just need to refine a few things, but it is basically done:

• the configure part
• understand/advise/workaround the fact that if in an input file you specify POSITION and DISTANCE both using the GPU you'll get a crash and I have NULL ideas why
• maybe adding a few more tests, just to be sure.
• understand why all the CI does not like my code

[Iximiel]

@gtribello I think it is done: with this last change I should have solved the cppcheck about __PLUMED_HAS_OPENACC not being defined somewhere

My idea is to activate __PLUMED_USE_OPENACC with __PLUMED_HAS_OPENACC while it is "experimental" and then remove these "preheaders" and use __PLUMED_HAS_OPENACC directly

Moreover now there is a "standard" way of compiling with openacc

I think now this is ready to be merged into final-backpropagation, so that I can rebase on these changes the work I am doing on the secondarystructure-2 branch