[Python | GPU] Reduce operation fails code generation on GPU

[FlorianDeconinck]

Tested only in DaCe v1

The following code should work

import dace
import cupy as cp
import numpy as np

@dace.program
def multidimred(a, b):
    b[:] = np.sum(a, axis=(0, 1))


a = cp.random.rand(2, 2, 6)
b = cp.random.rand(6)
multidimred(a, b)

but it dies in codegen:

  File "/home/fgdeconi/work/git/ndsl/external/dace/dace/codegen/dispatcher.py", line 466, in dispatch_scope
    self._map_dispatchers[map_schedule].generate_scope(sdfg, cfg, sub_dfg, state_id, function_stream,
  File "/home/fgdeconi/work/git/ndsl/external/dace/dace/codegen/targets/cuda.py", line 1656, in generate_scope
    cudastream = '__state->gpu_context->streams[%d]' % scope_entry._cuda_stream
                                                       ^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'MapEntry' object has no attribute '_cuda_stream'

There's a unit test testing this exact code. But the test doesn't call the program directly, instead it runs

    sdfg = multidimred.to_sdfg(a, b)
    sdfg.apply_gpu_transformations()

which, indeed fixes the present issue, but arguably shouldn't be needed.

PS: The cupy.max is not supported.