Merge origin/main

Author: tehrengruber

ci/cscs-ci.yml

@@ -19,7 +19,7 @@ stages:
 - test
 
 build py38 baseimage:
-  extends: .container-builder
+  extends: .container-builder-cscs-zen2
   stage: baseimage
   # we create a tag that depends on the SHA value of ci/base.Dockerfile, this way
   # a new base image is only built when the SHA of this file changes
@@ -52,7 +52,7 @@ build py310 baseimage:
     <<: *py310
 
 build py38 image:
-  extends: .container-builder
+  extends: .container-builder-cscs-zen2
   needs: ["build py38 baseimage"]
   stage: image
   variables:

src/gt4py/_core/definitions.py

@@ -444,6 +444,8 @@ def shape(self) -> tuple[int, ...]: ...
     @property
     def dtype(self) -> Any: ...
 
+    def item(self) -> Any: ...
+
     def astype(self, dtype: npt.DTypeLike) -> NDArrayObject: ...
 
     def __getitem__(self, item: Any) -> NDArrayObject: ...

src/gt4py/next/allocators.py

@@ -44,7 +44,7 @@
 CUPY_DEVICE: Final[Literal[None, core_defs.DeviceType.CUDA, core_defs.DeviceType.ROCM]] = (
     None
     if not cp
-    else (core_defs.DeviceType.ROCM if cp.cuda.get_hipcc_path() else core_defs.DeviceType.CUDA)
+    else (core_defs.DeviceType.ROCM if cp.cuda.runtime.is_hip else core_defs.DeviceType.CUDA)
 )
 
 

src/gt4py/next/common.py

@@ -623,14 +623,17 @@ def asnumpy(self) -> np.ndarray: ...
     def remap(self, index_field: ConnectivityField | fbuiltins.FieldOffset) -> Field: ...
 
     @abc.abstractmethod
-    def restrict(self, item: AnyIndexSpec) -> Field | core_defs.ScalarT: ...
+    def restrict(self, item: AnyIndexSpec) -> Field: ...
+
+    @abc.abstractmethod
+    def as_scalar(self) -> core_defs.ScalarT: ...
 
     # Operators
     @abc.abstractmethod
     def __call__(self, index_field: ConnectivityField | fbuiltins.FieldOffset) -> Field: ...
 
     @abc.abstractmethod
-    def __getitem__(self, item: AnyIndexSpec) -> Field | core_defs.ScalarT: ...
+    def __getitem__(self, item: AnyIndexSpec) -> Field: ...
 
     @abc.abstractmethod
     def __abs__(self) -> Field: ...
@@ -896,6 +899,9 @@ def ndarray(self) -> Never:
     def asnumpy(self) -> Never:
         raise NotImplementedError()
 
+    def as_scalar(self) -> Never:
+        raise NotImplementedError()
+
     @functools.cached_property
     def domain(self) -> Domain:
         return Domain(dims=(self.dimension,), ranges=(UnitRange.infinite(),))
@@ -947,9 +953,7 @@ def remap(self, index_field: ConnectivityField | fbuiltins.FieldOffset) -> Conne
 
     __call__ = remap
 
-    def restrict(self, index: AnyIndexSpec) -> core_defs.IntegralScalar:
-        if is_int_index(index):
-            return index + self.offset
+    def restrict(self, index: AnyIndexSpec) -> Never:
         raise NotImplementedError()  # we could possibly implement with a FunctionField, but we don't have a use-case
 
     __getitem__ = restrict

src/gt4py/next/embedded/nd_array_field.py

@@ -120,6 +120,13 @@ def asnumpy(self) -> np.ndarray:
         else:
             return np.asarray(self._ndarray)
 
+    def as_scalar(self) -> core_defs.ScalarT:
+        if self.domain.ndim != 0:
+            raise ValueError(
+                "'as_scalar' is only valid on 0-dimensional 'Field's, got a {self.domain.ndim}-dimensional 'Field'."
+            )
+        return self.ndarray.item()
+
     @property
     def codomain(self) -> type[core_defs.ScalarT]:
         return self.dtype.scalar_type
@@ -204,15 +211,11 @@ def remap(
 
     __call__ = remap  # type: ignore[assignment]
 
-    def restrict(self, index: common.AnyIndexSpec) -> common.Field | core_defs.ScalarT:
+    def restrict(self, index: common.AnyIndexSpec) -> common.Field:
         new_domain, buffer_slice = self._slice(index)
-
         new_buffer = self.ndarray[buffer_slice]
-        if len(new_domain) == 0:
-            # TODO: assert core_defs.is_scalar_type(new_buffer), new_buffer
-            return new_buffer  # type: ignore[return-value] # I don't think we can express that we return `ScalarT` here
-        else:
-            return self.__class__.from_array(new_buffer, domain=new_domain)
+        new_buffer = self.__class__.array_ns.asarray(new_buffer)
+        return self.__class__.from_array(new_buffer, domain=new_domain)
 
     __getitem__ = restrict
 
@@ -433,7 +436,7 @@ def inverse_image(
 
         return new_dims
 
-    def restrict(self, index: common.AnyIndexSpec) -> common.Field | core_defs.IntegralScalar:
+    def restrict(self, index: common.AnyIndexSpec) -> common.Field:
         cache_key = (id(self.ndarray), self.domain, index)
 
         if (restricted_connectivity := self._cache.get(cache_key, None)) is None:

src/gt4py/next/embedded/operators.py

@@ -187,8 +187,7 @@ def _tuple_at(
 ) -> core_defs.Scalar | tuple[core_defs.ScalarT | tuple, ...]:
     @utils.tree_map
     def impl(field: common.Field | core_defs.Scalar) -> core_defs.Scalar:
-        res = field[pos] if common.is_field(field) else field
-        res = res.item() if hasattr(res, "item") else res  # extract scalar value from array
+        res = field[pos].as_scalar() if common.is_field(field) else field
         assert core_defs.is_scalar_type(res)
         return res
 

src/gt4py/next/ffront/func_to_foast.py

@@ -289,7 +289,7 @@ def visit_Subscript(self, node: ast.Subscript, **kwargs) -> foast.Subscript:
             index = self._match_index(node.slice)
         except ValueError:
             raise errors.DSLError(
-                self.get_location(node.slice), "eXpected an integral index."
+                self.get_location(node.slice), "Expected an integral index."
             ) from None
 
         return foast.Subscript(

src/gt4py/next/ffront/past_passes/type_deduction.py

@@ -217,12 +217,12 @@ def visit_Call(self, node: past.Call, **kwargs):
                         f"'{new_kwargs['out'].type}'."
                     )
             elif new_func.id in ["minimum", "maximum"]:
-                if new_args[0].type != new_args[1].type:
+                if arg_types[0] != arg_types[1]:
                     raise ValueError(
                         f"First and second argument in '{new_func.id}' must be of the same type."
-                        f"Got '{new_args[0].type}' and '{new_args[1].type}'."
+                        f"Got '{arg_types[0]}' and '{arg_types[1]}'."
                     )
-                return_type = new_args[0].type
+                return_type = arg_types[0]
             else:
                 raise AssertionError(
                     "Only calls to 'FieldOperator', 'ScanOperator' or 'minimum' and 'maximum' builtins allowed."

src/gt4py/next/ffront/past_to_itir.py

@@ -332,7 +332,7 @@ def _visit_stencil_call_out_arg(
     ) -> tuple[itir.Expr, itir.FunCall]:
         if isinstance(out_arg, past.Subscript):
             # as the ITIR does not support slicing a field we have to do a deeper
-            #  inspection of the PAST to emulate the behaviour
+            # inspection of the PAST to emulate the behaviour
             out_field_name: past.Name = out_arg.value
             return (
                 self._construct_itir_out_arg(out_field_name),
@@ -409,12 +409,11 @@ def visit_BinOp(self, node: past.BinOp, **kwargs) -> itir.FunCall:
         )
 
     def visit_Call(self, node: past.Call, **kwargs) -> itir.FunCall:
-        if node.func.id in ["maximum", "minimum"] and len(node.args) == 2:
+        if node.func.id in ["maximum", "minimum"]:
+            assert len(node.args) == 2
             return itir.FunCall(
                 fun=itir.SymRef(id=node.func.id),
                 args=[self.visit(node.args[0]), self.visit(node.args[1])],
             )
         else:
-            raise AssertionError(
-                "Only 'minimum' and 'maximum' builtins supported supported currently."
-            )
+            raise NotImplementedError("Only 'minimum', and 'maximum' builtins supported currently.")

src/gt4py/next/iterator/embedded.py

@@ -919,7 +919,7 @@ def _translate_named_indices(
         return tuple(domain_slice)
 
     def field_getitem(self, named_indices: NamedFieldIndices) -> Any:
-        return self._ndarrayfield[self._translate_named_indices(named_indices)]
+        return self._ndarrayfield[self._translate_named_indices(named_indices)].as_scalar()
 
     def field_setitem(self, named_indices: NamedFieldIndices, value: Any):
         if common.is_mutable_field(self._ndarrayfield):
@@ -1040,6 +1040,7 @@ class IndexField(common.Field):
     """
 
     _dimension: common.Dimension
+    _cur_index: Optional[core_defs.IntegralScalar] = None
 
     @property
     def __gt_domain__(self) -> common.Domain:
@@ -1055,7 +1056,10 @@ def __gt_builtin_func__(func: Callable, /) -> NoReturn:  # type: ignore[override
 
     @property
     def domain(self) -> common.Domain:
-        return common.Domain((self._dimension, common.UnitRange.infinite()))
+        if self._cur_index is None:
+            return common.Domain((self._dimension, common.UnitRange.infinite()))
+        else:
+            return common.Domain()
 
     @property
     def codomain(self) -> type[core_defs.int32]:
@@ -1072,16 +1076,24 @@ def ndarray(self) -> core_defs.NDArrayObject:
     def asnumpy(self) -> np.ndarray:
         raise NotImplementedError()
 
+    def as_scalar(self) -> core_defs.IntegralScalar:
+        if self.domain.ndim != 0:
+            raise ValueError(
+                "'as_scalar' is only valid on 0-dimensional 'Field's, got a {self.domain.ndim}-dimensional 'Field'."
+            )
+        assert self._cur_index is not None
+        return self._cur_index
+
     def remap(self, index_field: common.ConnectivityField | fbuiltins.FieldOffset) -> common.Field:
         # TODO can be implemented by constructing and ndarray (but do we know of which kind?)
         raise NotImplementedError()
 
-    def restrict(self, item: common.AnyIndexSpec) -> common.Field | core_defs.int32:
+    def restrict(self, item: common.AnyIndexSpec) -> common.Field:
         if common.is_absolute_index_sequence(item) and all(common.is_named_index(e) for e in item):  # type: ignore[arg-type] # we don't want to pollute the typing of `is_absolute_index_sequence` for this temporary code # fmt: off
             d, r = item[0]
             assert d == self._dimension
-            assert isinstance(r, int)
-            return self.dtype.scalar_type(r)
+            assert isinstance(r, core_defs.INTEGRAL_TYPES)
+            return self.__class__(self._dimension, r)  # type: ignore[arg-type] # not sure why the assert above does not work
         # TODO set a domain...
         raise NotImplementedError()
 
@@ -1195,8 +1207,12 @@ def remap(self, index_field: common.ConnectivityField | fbuiltins.FieldOffset) -
         # TODO can be implemented by constructing and ndarray (but do we know of which kind?)
         raise NotImplementedError()
 
-    def restrict(self, item: common.AnyIndexSpec) -> common.Field | core_defs.ScalarT:
+    def restrict(self, item: common.AnyIndexSpec) -> common.Field:
         # TODO set a domain...
+        return self
+
+    def as_scalar(self) -> core_defs.ScalarT:
+        assert self.domain.ndim == 0
         return self._value
 
     __call__ = remap

tests/next_tests/integration_tests/feature_tests/ffront_tests/test_execution.py

@@ -321,6 +321,21 @@ def testee(inp: gtx.Field[[KDim], float]) -> gtx.Field[[KDim], float]:
     cases.verify(cartesian_case, testee, inp, out=out, ref=expected)
 
 
+def test_single_value_field(cartesian_case):
+    @gtx.field_operator
+    def testee_fo(a: cases.IKField) -> cases.IKField:
+        return a
+
+    @gtx.program
+    def testee_prog(a: cases.IKField):
+        testee_fo(a, out=a[1:2, 3:4])
+
+    a = cases.allocate(cartesian_case, testee_prog, "a")()
+    ref = a[1, 3]
+
+    cases.verify(cartesian_case, testee_prog, a, inout=a[1, 3], ref=ref)
+
+
 def test_astype_int(cartesian_case):  # noqa: F811 # fixtures
     @gtx.field_operator
     def testee(a: cases.IFloatField) -> gtx.Field[[IDim], int64]:

tests/next_tests/integration_tests/feature_tests/iterator_tests/test_horizontal_indirection.py

@@ -70,7 +70,7 @@ def test_simple_indirection(program_processor):
 
     ref = np.zeros(shape, dtype=inp.dtype)
     for i in range(shape[0]):
-        ref[i] = inp.ndarray[i + 1 - 1] if cond[i] < 0.0 else inp.ndarray[i + 1 + 1]
+        ref[i] = inp.asnumpy()[i + 1 - 1] if cond.asnumpy()[i] < 0.0 else inp.asnumpy()[i + 1 + 1]
 
     run_processor(
         conditional_indirection[cartesian_domain(named_range(IDim, 0, shape[0]))],
@@ -101,7 +101,7 @@ def test_direct_offset_for_indirection(program_processor):
 
     ref = np.zeros(shape)
     for i in range(shape[0]):
-        ref[i] = inp[i + cond[i]]
+        ref[i] = inp.asnumpy()[i + cond.asnumpy()[i]]
 
     run_processor(
         direct_indirection[cartesian_domain(named_range(IDim, 0, shape[0]))],

tests/next_tests/integration_tests/multi_feature_tests/iterator_tests/test_anton_toy.py

@@ -65,11 +65,15 @@ def fencil(x, y, z, out, inp):
 def naive_lap(inp):
     shape = [inp.shape[0] - 2, inp.shape[1] - 2, inp.shape[2]]
     out = np.zeros(shape)
+    inp_data = inp.asnumpy()
     for i in range(1, shape[0] + 1):
         for j in range(1, shape[1] + 1):
             for k in range(0, shape[2]):
-                out[i - 1, j - 1, k] = -4 * inp[i, j, k] + (
-                    inp[i + 1, j, k] + inp[i - 1, j, k] + inp[i, j + 1, k] + inp[i, j - 1, k]
+                out[i - 1, j - 1, k] = -4 * inp_data[i, j, k] + (
+                    inp_data[i + 1, j, k]
+                    + inp_data[i - 1, j, k]
+                    + inp_data[i, j + 1, k]
+                    + inp_data[i, j - 1, k]
                 )
     return out
 

tests/next_tests/unit_tests/embedded_tests/test_nd_array_field.py

@@ -468,10 +468,11 @@ def test_absolute_indexing_value_return():
     field = common._field(np.reshape(np.arange(100, dtype=np.int32), (10, 10)), domain=domain)
 
     named_index = ((IDim, 12), (JDim, 6))
+    assert common.is_field(field)
     value = field[named_index]
 
-    assert isinstance(value, np.int32)
-    assert value == 21
+    assert common.is_field(value)
+    assert value.as_scalar() == 21
 
 
 @pytest.mark.parametrize(
@@ -568,14 +569,17 @@ def test_relative_indexing_slice_3D(index, expected_shape, expected_domain):
 
 @pytest.mark.parametrize(
     "index, expected_value",
-    [((1, 0), 10), ((0, 1), 1)],
+    [
+        ((1, 0), 10),
+        ((0, 1), 1),
+    ],
 )
 def test_relative_indexing_value_return(index, expected_value):
     domain = common.Domain(dims=(IDim, JDim), ranges=(UnitRange(5, 15), UnitRange(2, 12)))
     field = common._field(np.reshape(np.arange(100, dtype=int), (10, 10)), domain=domain)
     indexed_field = field[index]
 
-    assert indexed_field == expected_value
+    assert indexed_field.as_scalar() == expected_value
 
 
 @pytest.mark.parametrize("lazy_slice", [lambda f: f[13], lambda f: f[:5, :3, :2]])