index_add

ivy/functional/frontends/torch/indexing_slicing_joining_mutating_ops.py

@@ -280,3 +280,45 @@ def where(condition, input=None, other=None):
 @to_ivy_arrays_and_back
 def conj(input):
     return ivy.conj(input)
+
+
+@to_ivy_arrays_and_back
+def index_add(input, dim, index, source, *, alpha=1, out=None):
+    # Potential Bug:
+    #   There is an issue with the torch backend (not caused by ivy)
+    #   where half precision (float16) values get ignored in summation:
+    #
+    #   >>> a = torch.tensor(-14., dtype=torch.float16)
+    #   >>> b = torch.tensor(1.014, dtype=torch.float16)
+    #   >>> a+b
+    #   tensor(-12.9844, dtype=torch.float16)
+    #   >>> a = torch.tensor(-24., dtype=torch.float16)
+    #   >>> a+b
+    #   tensor(-22.9844, dtype=torch.float16)
+    #   >>> a = torch.tensor(-34., dtype=torch.float16)
+    #   >>> a+b
+    #   tensor(-33., dtype=torch.float16)
+    #   >>>
+    input = ivy.swapaxes(input, dim, 0)
+    source = ivy.swapaxes(source, dim, 0)
+    _to_adds = []
+    index = sorted(zip(ivy.to_list(index), range(len(index))), key=(lambda x: x[0]))
+    while index:
+        _curr_idx = index[0][0]
+        while len(_to_adds) < _curr_idx:
+            _to_adds.append(ivy.zeros_like(source[0]))
+        _to_add_cum = ivy.get_item(source, index[0][1])
+        while (1 < len(index)) and (index[0][0] == index[1][0]):
+            _to_add_cum = ivy.add(_to_add_cum, ivy.get_item(source, index.pop(1)[1]))
+        index.pop(0)
+        _to_adds.append(_to_add_cum)
+    while len(_to_adds) < input.shape[0]:
+        _to_adds.append(ivy.zeros_like(source[0]))
+    _to_adds = ivy.stack(_to_adds)
+    if len(input.shape) < 2:
+        # Added this line due to the paddle backend treating scalars as 1-d arrays
+        _to_adds = ivy.flatten(_to_adds)
+
+    ret = ivy.add(input, _to_adds, alpha=alpha)
+    ret = ivy.swapaxes(ret, 0, dim, out=out)
+    return ret

ivy_tests/test_ivy/test_frontends/test_torch/test_indexing_slicing_joining_mutating_ops.py

@@ -1217,3 +1217,94 @@ def test_torch_conj(
         on_device=on_device,
         input=x[0],
     )
+
+
+@st.composite
+def _arrays_dim_idx_n_dtypes(draw):
+    num_dims = draw(st.shared(helpers.ints(min_value=1, max_value=4), key="num_dims"))
+    num_arrays = 2
+    common_shape = draw(
+        helpers.lists(
+            x=helpers.ints(min_value=2, max_value=3),
+            min_size=num_dims - 1,
+            max_size=num_dims - 1,
+        )
+    )
+    _dim = draw(helpers.ints(min_value=0, max_value=num_dims - 1))
+    unique_dims = draw(
+        helpers.lists(
+            x=helpers.ints(min_value=2, max_value=3),
+            min_size=num_arrays,
+            max_size=num_arrays,
+        )
+    )
+
+    min_dim = min(unique_dims)
+    max_dim = max(unique_dims)
+    _idx = draw(
+        helpers.array_values(
+            shape=min_dim,
+            dtype="int64",
+            min_value=0,
+            max_value=max_dim,
+            exclude_min=False,
+        )
+    )
+
+    xs = list()
+    available_input_types = draw(helpers.get_dtypes("numeric"))
+    available_input_types.remove("float16")  # half summation unstable in backends
+    input_dtypes = draw(
+        helpers.array_dtypes(
+            available_dtypes=available_input_types,
+            num_arrays=num_arrays,
+            shared_dtype=True,
+        )
+    )
+    for ud, dt in zip(unique_dims, input_dtypes):
+        x = draw(
+            helpers.array_values(
+                shape=common_shape[:_dim] + [ud] + common_shape[_dim:],
+                dtype=dt,
+                large_abs_safety_factor=2.5,
+                small_abs_safety_factor=2.5,
+                safety_factor_scale="log",
+            )
+        )
+        xs.append(x)
+    return xs, input_dtypes, _dim, _idx
+
+
+# index_add
+@handle_frontend_test(
+    fn_tree="torch.index_add",
+    xs_dtypes_dim_idx=_arrays_dim_idx_n_dtypes(),
+    alpha=st.integers(min_value=1, max_value=2),
+)
+def test_torch_index_add(
+    *,
+    xs_dtypes_dim_idx,
+    alpha,
+    on_device,
+    fn_tree,
+    frontend,
+    test_flags,
+):
+    xs, input_dtypes, axis, indices = xs_dtypes_dim_idx
+    if xs[0].shape[axis] < xs[1].shape[axis]:
+        source, input = xs
+    else:
+        input, source = xs
+    helpers.test_frontend_function(
+        input_dtypes=[input_dtypes[0], "int64", input_dtypes[1]],
+        frontend=frontend,
+        test_flags=test_flags,
+        fn_tree=fn_tree,
+        on_device=on_device,
+        rtol=1e-03,
+        input=input,
+        dim=axis,
+        index=indices,
+        source=source,
+        alpha=alpha,
+    )