Enabling Globals for Unstructured Backends

[mroethlin]

Technical Description

This PR enables globals for the unstructured backends. Furthermore, an unreported issue where globals were not propagated from the wrapper class to the stencil class was fixed. Additionally, a bug in the unstructured cuda codegen was fixed when translating stencils that only use dense dimensions.

Resolves / Enhances

Fixes #1030
Fixes #1028

Notes

The methods to set and get globals in the cuda backend are on the inner stencils. This will be addressed in this issue. Also, a method to communicate globals from FORTRAN will need to be devised (not addressed yet).

Testing

New tests in dawn4py and a new unstructured integration test to test the correct operation of the CXXNaiveIco backend. CudaIco backend tested manually.

Dependencies

This PR is independent.

[mroethlin]

launch jenkins

[mroethlin]

launch jenkins

[mroethlin]

launch jenkins

[Stagno]

There is also another problem, if you gtclang this (-backend=c++-naive)

#include "gtclang_dsl_defs/gtclang_dsl.hpp"
using namespace gtclang::dsl;

globals {
  double myval;
};

stencil Test {
  storage field_a, field_b;

  Do {
  	myval = 1.0;
  	vertical_region(k_start, k_end) {
	    field_a = field_b + myval;
    }
  	myval = 2.0;
    vertical_region(k_start, k_end) {
	    field_b = field_a + myval;
    }
  }
};

you'll see that setting a global via "control flow" will bypass the setter call and directly update m_globals of the wrapper class. I think the only "safe" option is having a const globals& in the inner class (in each backend) referencing a single globals object present in the wrapper class.

[mroethlin]

launch jenkins

[mroethlin]

launch jenkins

[mroethlin]

launch jenkins