mpi: Generate deterministic code for overlap mode

[georgebisbas]

Overlap mode did not have deterministic code generation.

static void remainder0(struct dataobj *restrict u_vec, const float a, const float dt, float r0, float r1, float r2, int t0, int t1, const int x_M, const int x_m, const int y_M, const int y_m, const int nthreads)
{
  compute0(a,u_vec,dt,r0,r1,r2,t0,t1,x_M - 2,x_m + 2,MIN(y_M, y_m + 1),y_m,nthreads);
  compute0(a,u_vec,dt,r0,r1,r2,t0,t1,x_M - 2,x_m + 2,y_M,MAX(y_m, y_M - 1),nthreads);
  compute0(a,u_vec,dt,r0,r1,r2,t0,t1,MIN(x_M, x_m + 1),x_m,y_M - 2,y_m + 2,nthreads);
  compute0(a,u_vec,dt,r0,r1,r2,t0,t1,MIN(x_M, x_m + 1),x_m,MIN(y_M, y_m + 1),y_m,nthreads);
  compute0(a,u_vec,dt,r0,r1,r2,t0,t1,MIN(x_M, x_m + 1),x_m,y_M,MAX(y_m, y_M - 1),nthreads);
  compute0(a,u_vec,dt,r0,r1,r2,t0,t1,x_M,MAX(x_m, x_M - 1),y_M - 2,y_m + 2,nthreads);
  compute0(a,u_vec,dt,r0,r1,r2,t0,t1,x_M,MAX(x_m, x_M - 1),MIN(y_M, y_m + 1),y_m,nthreads);
  compute0(a,u_vec,dt,r0,r1,r2,t0,t1,x_M,MAX(x_m, x_M - 1),y_M,MAX(y_m, y_M - 1),nthreads);
}

compute0 calls could have random order.

Also, MIN/MAX macros could have random order. (this is a side-effect, not overlap specific)

To reproduce, you could run a script multiple times with DEVITO_MPI=overlap

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (313c980) 86.69% compared to head (a272ba3) 86.69%.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #2303   +/-   ##
=======================================
  Coverage   86.69%   86.69%           
=======================================
  Files         229      229           
  Lines       43043    43043           
  Branches     7983     7983           
=======================================
  Hits        37318    37318           
  Misses       5034     5034           
  Partials      691      691           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mloubout]

same filter_sorted

[georgebisbas]

For this I think filter-sorted is redundant. What would it offer more?

[FabioLuporini]

I think this is conceptually wrong and useless:

Conceptually wrong because if all passes appending headers had to ensure uniquess it would be painful. BTW it's not even just the headers, it would be everything that can get passed to metadata.

Which brings us to the second point, useless: see here:
https://github.com/devitocodes/devito/blob/master/devito/passes/iet/engine.py#L123

[FabioLuporini]

obviously if you instead added it because it generates bad code, then there's a bug somewhere else

[georgebisbas]

Correct! Thanks, it was a side-effect of the compute regions non-determinism, catching the first (min/max) area to be computed

[georgebisbas]

@mloubout @FabioLuporini this should be ok?

[FabioLuporini]

I think this is conceptually wrong and useless:

Conceptually wrong because if all passes appending headers had to ensure uniquess it would be painful. BTW it's not even just the headers, it would be everything that can get passed to metadata.

Which brings us to the second point, useless: see here:
https://github.com/devitocodes/devito/blob/master/devito/passes/iet/engine.py#L123

[FabioLuporini]

Thanks, merged