Corrected result.data implementation. (quantumlib#5153)

When looking into quantumlib#4774 auto type promotion to objects raised a ValueError and was not tested. This PR fixes this and introduces the intended behavior as well as boosting performance slightly when qubit counts are < 64.

cirq/study/result.py

@@ -142,14 +142,16 @@ def dataframe_from_measurements(measurements: Mapping[str, np.ndarray]) -> pd.Da
         """
         # Convert to a DataFrame with columns as measurement keys, rows as
         # repetitions and a big endian integer for individual measurements.
-        converted_dict = {
-            key: [value.big_endian_bits_to_int(m_vals) for m_vals in val]
-            for key, val in measurements.items()
-        }
-        # Note that when a numpy array is produced from this data frame,
-        # Pandas will try to use np.int64 as dtype, but will upgrade to
-        # object if any value is too large to fit.
-        return pd.DataFrame(converted_dict, dtype=np.int64)
+        converted_dict = {}
+        for key, bitstrings in measurements.items():
+            _, n = bitstrings.shape
+            dtype = object if n > 63 else np.int64
+            basis = 2 ** np.arange(n, dtype=dtype)[::-1]
+            converted_dict[key] = np.sum(basis * bitstrings, axis=1)
+
+        # Use objects to accomodate more than 64 qubits if needed.
+        dtype = object if any(bs.shape[1] > 63 for _, bs in measurements.items()) else np.int64
+        return pd.DataFrame(converted_dict, dtype=dtype)
 
     @staticmethod
     @deprecated(

cirq/study/result_test.py

@@ -156,7 +156,9 @@ def test_df():
             'c': np.array([[0], [0], [1], [0], [1]], dtype=bool),
         },
     )
-    remove_end_measurements = pd.DataFrame(data={'ab': [1, 1, 2], 'c': [0, 1, 0]}, index=[1, 2, 3])
+    remove_end_measurements = pd.DataFrame(
+        data={'ab': [1, 1, 2], 'c': [0, 1, 0]}, index=[1, 2, 3], dtype=np.int64
+    )
 
     pd.testing.assert_frame_equal(result.data.iloc[1:-1], remove_end_measurements)
 
@@ -166,6 +168,21 @@ def test_df():
     assert df.c.value_counts().to_dict() == {0: 3, 1: 2}
 
 
+def test_df_large():
+    result = cirq.ResultDict(
+        params=cirq.ParamResolver({}),
+        measurements={
+            'a': np.array([[0 for _ in range(76)]] * 10_000, dtype=bool),
+            'd': np.array([[1 for _ in range(76)]] * 10_000, dtype=bool),
+        },
+    )
+
+    assert np.all(result.data['a'] == 0)
+    assert np.all(result.data['d'] == 0xFFF_FFFFFFFF_FFFFFFFF)
+    assert result.data['a'].dtype == object
+    assert result.data['d'].dtype == object
+
+
 def test_histogram():
     result = cirq.ResultDict(
         params=cirq.ParamResolver({}),