Tech slk

[andrea-ferretti]

This branch contains several developments:

• implementation of g-parallelism for SE (switched off for SC and RT runs since not currently compatible)
• cleanup of QP_ppa_cohsex (at least for what concern GPU porting)
• cleanup of other GPU-oriented operations (notably including reduction of the GPU memory footprint)
• cleanup and fixes to the MPA driver for the self-energy
• minor fixes

[sangallidavide]

Just one comment for eval_G_minus_G

I see you replaced the G_m_G_maxval_tmp procedure with the call to G_m_G_maxval=maxval(G_m_G)

However, this is done before the PP_redux(G_m_G), where G_m_G might be distributed of the MPI tasks. This would result in different MPI tasks having different values.

One should either move the maxval after the PP_Redux or (maybe more efficient ?) code

allocate(G_m_G_maxval_tmp(ncpu)
G_m_G_maxval_tmp(myid)=maxval(G_m_G)
PP_redux(G_m_G_maxval_tmp)
G_m_G_maxval=maxval(G_m_G_maxval_tmp)

[andrea-ferretti]

Absolutely, this needs fixing (I forgot to do it). I'll implement the second option with the redux (nicer not to have to do reductions with the max function).

[sangallidavide]

Rethinking. Rather than the second option, which anyway would do a redux on the whole G_m_G, I'd move back to the previous structure

 YAMBO_ALLOC(G_m_G_maxval_tmp,(iGo_))                                   
 G_m_G_maxval_tmp=0                                                    

with later

do i2=1,oGo_
     ..
     G_m_G_maxval_tmp(i2)=maxval(G_m_G(:,i2))       
enddo
PP_redux(G_m_G_maxval_tmp)
G_m_G_maxval=maxval(G_m_G_maxval_tmp)

[andrea-ferretti]

I agree, this is better ! coded and under testing now