Make prod work with u32 default type (#58)

Quantco · Aug 22, 2024 · 9837b83 · 9837b83
1 parent af78c47
commit 9837b83
Show file tree

Hide file tree

Showing 2 changed files with 39 additions and 7 deletions.
diff --git a/ndonnx/_core/_numericimpl.py b/ndonnx/_core/_numericimpl.py
@@ -663,7 +663,7 @@ def prod(
         else:
             axes = axis  # type: ignore
 
-        x = x.astype(_determine_reduce_op_dtype(x, dtype))
+        x = x.astype(_determine_reduce_op_dtype(x, dtype, dtypes.uint32))
 
         if isinstance(x.dtype, dtypes.NullableNumerical):
             fill_value = ndx.asarray(1, dtype=x.dtype.values)
@@ -760,7 +760,7 @@ def sum(
         else:
             axes = axis  # type: ignore
 
-        x = x.astype(_determine_reduce_op_dtype(x, dtype))
+        x = x.astype(_determine_reduce_op_dtype(x, dtype, dtypes.uint64))
 
         # Fill any nulls with 0
         if isinstance(x.dtype, dtypes.NullableNumerical):
@@ -972,14 +972,22 @@ def _via_i64_f64(
 
 
 def _determine_reduce_op_dtype(
-    x: Array, dtype: dtypes.CoreType | dtypes.StructType | None
+    x: Array,
+    dtype: dtypes.CoreType | dtypes.StructType | None,
+    maximum_unsigned_dtype: dtypes.CoreType,
 ) -> dtypes.CoreType | dtypes.StructType:
     if dtype is not None:
         return dtype
     elif isinstance(x.dtype, dtypes.Unsigned):
-        return dtypes.uint64
+        if ndx.iinfo(x.dtype).bits <= ndx.iinfo(maximum_unsigned_dtype).bits:
+            return maximum_unsigned_dtype
+        else:
+            raise TypeError(f"Cannot reduce {x.dtype} to a smaller unsigned dtype")
     elif isinstance(x.dtype, dtypes.NullableUnsigned):
-        return dtypes.nuint64
+        if ndx.iinfo(x.dtype).bits <= ndx.iinfo(maximum_unsigned_dtype).bits:
+            return dtypes.promote_nullable(maximum_unsigned_dtype)
+        else:
+            raise TypeError(f"Cannot reduce {x.dtype} to a smaller unsigned dtype")
     elif isinstance(x.dtype, dtypes.Integral):
         return dtypes.int64
     elif isinstance(x.dtype, dtypes.NullableIntegral):

diff --git a/tests/test_core.py b/tests/test_core.py
@@ -633,14 +633,38 @@ def test_prod(dtype):
     assert_array_equal(y.to_numpy(), actual)
 
 
+@pytest.mark.parametrize(
+    "dtype",
+    [
+        ndx.uint8,
+        ndx.uint16,
+        ndx.uint32,
+        ndx.nuint8,
+        ndx.nuint16,
+        ndx.nuint32,
+    ],
+)
+def test_prod_unsigned(dtype):
+    # We intentionally deviate from the Array API standard and reduce for <= uint32
+    # using uint32 as our default unsigned type due to lack of kernel support for uint64
+    x = ndx.asarray([2, 2]).astype(dtype)
+    y = ndx.prod(x)
+    if isinstance(dtype, ndx.Nullable):
+        input = np.asarray([2, 2], dtype=dtype.values.to_numpy_dtype())
+        input = np.ma.masked_array(input, mask=False)
+    else:
+        input = np.asarray([2, 2], dtype=dtype.to_numpy_dtype())
+    actual = np.prod(input).astype(np.uint32)
+
+    assert_array_equal(y.to_numpy(), actual)
+
+
 @pytest.mark.parametrize(
     "dtype",
     [
         ndx.float64,
         ndx.nfloat64,
         ndx.uint64,
-        ndx.uint32,
-        ndx.uint8,
         ndx.nuint64,
     ],
 )