mpi: Instrument compute0 core after specialising as ComputeCall

[georgebisbas]

This PR aims to make compute0 "more special" for instrumenting

and then adds a timer for the compute0 core area in FULL mpi mode (for advanced2 profiling)

e.g now we get:

Global performance: [OI=6.55, 15.50 GFlops/s, 0.19 GPts/s]
Local performance:
  * section0[rank0]<21,50,50,8,8,40> ran in 0.72 s [OI=6.55, 15.56 GFlops/s, 0.19 GPts/s]
    + haloupdate0 ran in 0.00 s [0.01%]
    + halowait0 ran in 0.00 s [0.01%]
    + remainder0 ran in 0.55 s [76.68%]
    + compute0 ran in 0.17 s [23.32%]

Generated code-diff:

*** 42,48 ****
    double haloupdate0;
    double halowait0;
    double remainder0;
-   double compute0;
  } ;
  
  struct region0
--- 42,47 ----
***************
*** 81,89 ****
      START_TIMER(haloupdate0)
      haloupdate0(u_vec,comm,msg0,6,t0,nthreads);
      STOP_TIMER(haloupdate0,timers)
-     START_TIMER(compute0)
      compute0(a,msg0,u_vec,dt,r0,r1,r2,r3,t0,t1,x0_blk0_size,x_M - 16,x_m + 16,y0_blk0_size,y_M - 16,y_m + 16,z_M - 16,z_m + 16,nthreads);
-     STOP_TIMER(compute0,timers)
      START_TIMER(halowait0)
      halowait0(u_vec,t0,msg0,6,nthreads);
      STOP_TIMER(halowait0,timers)
--- 80,86 ----

[codecov]

Codecov Report

Merging #2143 (cf1ba20) into master (772243b) will increase coverage by 0.02%.
The diff coverage is 97.56%.

@@            Coverage Diff             @@
##           master    #2143      +/-   ##
==========================================
+ Coverage   87.09%   87.11%   +0.02%     
==========================================
  Files         223      223              
  Lines       39871    39886      +15     
  Branches     5170     5170              
==========================================
+ Hits        34726    34748      +22     
+ Misses       4568     4562       -6     
+ Partials      577      576       -1     

Impacted Files	Coverage Δ
devito/ir/iet/efunc.py	78.08% <95.23%> (+0.30%)	⬆️
devito/mpi/routines.py	93.22% <100.00%> (+0.03%)	⬆️
devito/operator/profiling.py	69.71% <100.00%> (+0.94%)	⬆️
devito/passes/iet/instrument.py	80.00% <100.00%> (+7.27%)	⬆️
tests/test_mpi.py	98.86% <100.00%> (+<0.01%)	⬆️

[FabioLuporini]

This doesn't scale -- if you had 12 different kinds of Callables, what would you do? 12 different routines, one for each of them?

Here's an alternative

In ir/iet/efuncs.py, alter:

class ElementalFunction(...):

    _Call_cls = ElementalCall

    ...

    def make_call(...):
          return self._Call_cls(...)

Also, the fact that your make_compute_func is essentially a copy-paste of make_efunc should have smelled bad

You can extend make_efunc s.t.

def make_efunc(..., efunc_type=None):
    ...

so eventually here in routines.py all you will need to do is:

class ComputeCall(ElementalCall):
    pass

class ComputeFunction(ElementalFunction):
    _Call_cls = ComputeCall

Then calling make_efunc(..., efunc_type=ComputeFunction)

I'm not claiming this is exceptionally clean (the whole make_efunc thing seems like an overkill to me, but what can we do, it's been around for ages...), but at least it scales

[georgebisbas]

Requested changes have ben addressed

[FabioLuporini]

This LGTG to me now. @mloubout ?