Connector does not infer correct type + wrong index calculation

[aydogdub]

Description

When applying a pass that moves transient arrays stored in GPU_Global out of a kernel, two bugs occur:

1. The added input connectors fail to infer the correct type and default to "void". This results in references to void, which are invalid in C++ and cause compilation errors.

Example: the generated code attempts to create a reference to void, which is not allowed when no explicit types for the connectors are provided.

leading to:

2. To work around (1), I tried explicitly setting the type for added connectors as pointer(data_descriptor.dtype), since the connector should be a reference to an array. While this compiles, it introduces a new bug: the array index is calculated incorrectly.

How I explicitly set connector types:

Now, the accessed value in mask_r12[i,j] is wrong.

How the generated code computes the index:

The correct generated code should compute the index:

mask_r12[i * N + j]

To Reproduce
Steps to reproduce the behavior:

1. Checkout to the branch newgpucodegen

2. Run the GPU test example provided in
   dace/tests/npbench/misc/azimint_naive_test.py, which triggers the error.

   • Currently, the connectors are set explicitly in the pass:
     dace/transformations/passes/move_array_out_of_kernel.py
     → This triggers bug (2).

3. To trigger bug (1) instead, search for "dtypes.pointer" in the code (i.e. dace/transformations/passes/move_array_out_of_kernel.py ) and remove the additional explicit input from any connector.

   • This will cause the auto-detection to fail and default to void.