Add stack as a callback (#1482)

exla/lib/exla/backend.ex

@@ -243,6 +243,17 @@ defmodule EXLA.Backend do
     jit([], expr_fun, tensors, [List.to_tuple(tensors)])
   end
 
+  @impl true
+  def stack(out, tensors, axis) do
+    out = Nx.to_template(out)
+
+    expr_fun = fn tensors ->
+      Nx.Defn.Expr.stack(out, Tuple.to_list(tensors), axis)
+    end
+
+    jit([], expr_fun, tensors, [List.to_tuple(tensors)])
+  end
+
   @impl true
   def slice(out, tensor, start_indices, lengths, strides) do
     out = Nx.to_template(out)

exla/lib/exla/defn.ex

@@ -1293,10 +1293,13 @@ defmodule EXLA.Defn do
   end
 
   defp to_operator(:concatenate, [[%Value{} | _rest] = tensors, axis], ans, _state) do
-    tensors =
-      tensors
-      |> Enum.map(&to_type(&1, ans.type))
+    tensors = Enum.map(tensors, &to_type(&1, ans.type))
+    Value.concatenate(tensors, axis, expr_to_typespec(ans))
+  end
 
+  defp to_operator(:stack, [[%Value{} | _rest] = tensors, axis], ans, _state) do
+    reshape_typespec = Typespec.tensor(ans.type, put_elem(ans.shape, axis, 1))
+    tensors = Enum.map(tensors, &(&1 |> to_type(ans.type) |> Value.reshape(reshape_typespec)))
     Value.concatenate(tensors, axis, expr_to_typespec(ans))
   end
 

nx/lib/nx.ex

@@ -3251,20 +3251,9 @@ defmodule Nx do
   """
   @doc type: :shape, from_backend: false
   def new_axis(tensor, axis, name \\ nil) when is_integer(axis) do
-    apply_vectorized(tensor, fn tensor, offset ->
-      %{shape: shape, names: names} = tensor = to_tensor(tensor)
-      rank = tuple_size(shape)
-      norm = if axis < 0, do: axis + rank + 1, else: axis + offset
-
-      if norm not in offset..tuple_size(shape) do
-        raise ArgumentError,
-              "new axis position for shape #{inspect(shape)} must be " <>
-                "a number between #{-rank - 1 + offset} and #{rank - offset}, got: #{axis}"
-      end
-
-      new_shape = Tuple.insert_at(shape, norm, 1)
-      new_names = List.insert_at(names, norm, name)
-      impl!(tensor).reshape(%{tensor | shape: new_shape, names: new_names}, tensor)
+    apply_vectorized(tensor, fn %{shape: shape, names: names} = tensor, offset ->
+      {shape, names, _axis} = Nx.Shape.new_axis(shape, names, axis, name, 1, offset)
+      impl!(tensor).reshape(%{tensor | shape: shape, names: names}, tensor)
     end)
   end
 
@@ -14668,28 +14657,35 @@ defmodule Nx do
         t
 
       [_ | _] = tensors ->
-        [%T{vectorized_axes: vectorized_axes} | _] =
-          tensors = broadcast_vectors(tensors, align_ranks: true)
+        concatenate_or_stack(
+          tensors,
+          fn shapes, names, offset -> Nx.Shape.concatenate(shapes, names, axis, offset) end,
+          fn out, tensors, axis -> list_impl!(tensors).concatenate(out, tensors, axis) end
+        )
+    end
+  end
 
-        offset = length(vectorized_axes)
-        tensors = if vectorized_axes != [], do: Enum.map(tensors, &devectorize/1), else: tensors
+  defp concatenate_or_stack(tensors, shape_and_name, callback) do
+    [%T{vectorized_axes: vectorized_axes} | _] =
+      tensors = broadcast_vectors(tensors, align_ranks: true)
 
-        {types, [s1 | _] = shapes, [n1 | _] = names} =
-          Enum.reduce(tensors, {[], [], []}, fn
-            %T{type: t, shape: s, names: n}, {types, shapes, names} ->
-              {[t | types], [s | shapes], [n | names]}
-          end)
+    offset = length(vectorized_axes)
+    tensors = if vectorized_axes != [], do: Enum.map(tensors, &devectorize/1), else: tensors
+
+    {types, shapes, names} =
+      Enum.reduce(tensors, {[], [], []}, fn
+        %T{type: t, shape: s, names: n}, {types, shapes, names} ->
+          {[t | types], [s | shapes], [n | names]}
+      end)
 
-        axis = Nx.Shape.normalize_axis(s1, axis, n1, offset)
-        output_type = Enum.reduce(types, &Nx.Type.merge/2)
+    output_type = Enum.reduce(types, &Nx.Type.merge/2)
 
-        {output_shape, output_names} =
-          Nx.Shape.concatenate(Enum.reverse(shapes), Enum.reverse(names), axis)
+    {output_shape, output_names, axis} =
+      shape_and_name.(Enum.reverse(shapes), Enum.reverse(names), offset)
 
-        out = %{hd(tensors) | type: output_type, shape: output_shape, names: output_names}
-        result = list_impl!(tensors).concatenate(out, tensors, axis)
-        vectorize(result, vectorized_axes)
-    end
+    out = %{hd(tensors) | type: output_type, shape: output_shape, names: output_names}
+    result = callback.(out, tensors, axis)
+    vectorize(result, vectorized_axes)
   end
 
   defp flatten_list_or_container(list) when is_list(list) do
@@ -14807,16 +14803,26 @@ defmodule Nx do
       >
 
   """
-  @doc type: :ndim, from_backend: false
+  @doc type: :ndim
   def stack(tensors, opts \\ []) do
     opts = keyword!(opts, axis: 0, name: nil)
     axis = opts[:axis]
     name = opts[:name]
 
-    tensors
-    |> flatten_list_or_container()
-    |> Enum.map(&Nx.new_axis(&1, axis, name))
-    |> Nx.concatenate(axis: axis)
+    case flatten_list_or_container(tensors) do
+      [] ->
+        raise ArgumentError, "no tensors were given to stack"
+
+      [t] ->
+        Nx.new_axis(t, axis, name)
+
+      [_ | _] = tensors ->
+        concatenate_or_stack(
+          tensors,
+          fn shapes, names, offset -> Nx.Shape.stack(shapes, names, axis, name, offset) end,
+          fn out, tensors, axis -> list_impl!(tensors).stack(out, tensors, axis) end
+        )
+    end
   end
 
   @doc """

nx/lib/nx/backend.ex

@@ -75,6 +75,7 @@ defmodule Nx.Backend do
   @callback put_slice(out :: tensor, tensor, tensor, list) :: tensor
   @callback gather(out :: tensor, input :: tensor, indices :: tensor, keyword) :: tensor
   @callback concatenate(out :: tensor, tensor, axis) :: tensor
+  @callback stack(out :: tensor, tensor, axis) :: tensor
   @callback select(out :: tensor, tensor, tensor, tensor) :: tensor
 
   @callback conv(out :: tensor, tensor, kernel :: tensor, keyword) :: tensor

nx/lib/nx/binary_backend.ex

@@ -1999,6 +1999,19 @@ defmodule Nx.BinaryBackend do
     offset
   end
 
+  @impl true
+  def stack(out, tensors, axis) do
+    %{shape: output_shape, type: {_, size} = output_type} = out
+
+    tensors
+    |> Enum.map(fn %{shape: shape} = t ->
+      t = as_type(%{t | type: output_type}, t)
+      {to_binary(t), Tuple.insert_at(shape, axis, 1)}
+    end)
+    |> bin_concatenate(size, axis, output_shape)
+    |> then(&from_binary(out, &1))
+  end
+
   @impl true
   def concatenate(out, tensors, axis) do
     %{shape: output_shape, type: {_, size} = output_type} = out

nx/lib/nx/defn/evaluator.ex

@@ -20,7 +20,7 @@ defmodule Nx.Defn.Evaluator do
   alias Nx.Defn.{Composite, Expr, Tree}
 
   @creation_ops [:eye, :iota, :from_binary]
-  @list_ops [:concatenate]
+  @list_ops [:concatenate, :stack]
   @indices_ops [:slice, :put_slice]
 
   @impl true

nx/lib/nx/defn/expr.ex

@@ -1201,6 +1201,12 @@ defmodule Nx.Defn.Expr do
     expr(out, context, :concatenate, [tensors, axis])
   end
 
+  @impl true
+  def stack(out, tensors, axis) do
+    {tensors, context} = to_exprs(tensors)
+    expr(out, context, :stack, [tensors, axis])
+  end
+
   @impl true
   def triangular_solve(out, a, b, opts) do
     {[a, b], context} = to_exprs([a, b])

nx/lib/nx/defn/grad.ex

@@ -614,6 +614,21 @@ defmodule Nx.Defn.Grad do
     [{x, g}]
   end
 
+  defp grad(:stack, [tensors, axis], ans, g) do
+    zero_axes = List.duplicate(0, Nx.rank(ans))
+    ans_shape_list = Tuple.to_list(ans.shape)
+
+    {pairs, _} =
+      Enum.map_reduce(tensors, 0, fn t, limit ->
+        current_limit = 1 + limit
+        start = List.replace_at(zero_axes, axis, limit)
+        len = List.replace_at(ans_shape_list, axis, 1)
+        {{t, Nx.slice(g, start, len)}, current_limit}
+      end)
+
+    pairs
+  end
+
   defp grad(:concatenate, [tensors, axis], ans, g) do
     zero_axes = List.duplicate(0, Nx.rank(ans))
     ans_shape_list = Tuple.to_list(ans.shape)

nx/lib/nx/defn/tree.ex

@@ -202,7 +202,8 @@ defmodule Nx.Defn.Tree do
     {[%{token | hooks: hooks}], acc}
   end
 
-  def apply_args(%T{data: %Expr{op: :concatenate, args: [list | args]}}, _type, acc, fun) do
+  def apply_args(%T{data: %Expr{op: op, args: [list | args]}}, _type, acc, fun)
+      when op in [:concatenate, :stack] do
     {list, acc} = Enum.map_reduce(list, acc, fun)
     {[list | args], acc}
   end

nx/lib/nx/shape.ex

@@ -1648,17 +1648,65 @@ defmodule Nx.Shape do
     {shape, names}
   end
 
+  @doc """
+  Returns the shape and name of new axis.
+  """
+  def new_axis(shape, names, axis, name, size, offset) do
+    rank = tuple_size(shape)
+    norm = if axis < 0, do: axis + rank + 1, else: axis + offset
+
+    if norm not in offset..tuple_size(shape) do
+      raise ArgumentError,
+            "new axis position for shape #{inspect(shape)} must be " <>
+              "a number between #{-rank - 1 + offset} and #{rank - offset}, got: #{axis}"
+    end
+
+    new_shape = Tuple.insert_at(shape, norm, size)
+    new_names = List.insert_at(names, norm, name)
+    {new_shape, new_names, norm}
+  end
+
+  @doc """
+  Returns the shape and names after a stack.
+
+  ## Examples
+
+      iex> Nx.Shape.stack([{3, 2}, {3, 2}, {3, 2}], [[nil, nil], [nil, :z], [:y, nil]], 0, :x, 0)
+      {{3, 3, 2}, [:x, :y, :z], 0}
+  """
+  def stack(shapes, names, axis, name, offset) do
+    names =
+      Enum.zip_with(names, fn zipped ->
+        Enum.reduce(zipped, &merge_names!(&1, &2, axis, axis))
+      end)
+
+    case Enum.uniq(shapes) do
+      [shape] ->
+        new_axis(shape, names, axis, name, length(shapes), offset)
+
+      shapes ->
+        raise ArgumentError,
+              "can only stack tensors of the same shape, got distinct shapes: #{inspect(shapes)}"
+    end
+  end
+
   @doc """
   Returns the shape and names after a concat.
 
   ## Examples
 
-      iex> Nx.Shape.concatenate([{2, 3, 2}, {1, 3, 2}, {4, 3, 2}], [[:x, :y, :z], [:x, :y, :z], [:x, :y, :z]], 0)
-      {{7, 3, 2}, [:x, :y, :z]}
+      iex> Nx.Shape.concatenate([{2, 3, 2}, {1, 3, 2}, {4, 3, 2}], [[:x, :y, :z], [:x, :y, :z], [:x, :y, :z]], 0, 0)
+      {{7, 3, 2}, [:x, :y, :z], 0}
   """
-  def concatenate(shapes, names, axis) do
-    names = validate_concat_names!(names, axis)
-    {concat_dims(shapes, axis), names}
+  def concatenate([s1 | _] = shapes, [n1 | _] = names, axis, offset) do
+    axis = normalize_axis(s1, axis, n1, offset)
+
+    names =
+      Enum.zip_with(names, fn zipped ->
+        Enum.reduce(zipped, &merge_names!(&1, &2, axis, axis))
+      end)
+
+    {concat_dims(shapes, axis), names, axis}
   end
 
   defp concat_dims([s1 | shapes] = all_shapes, axis) do
@@ -2120,15 +2168,6 @@ defmodule Nx.Shape do
     )
   end
 
-  defp validate_concat_names!(names, axis) do
-    _ =
-      Enum.zip_with(names, fn zipped ->
-        Enum.reduce(zipped, &merge_names!(&1, &2, axis, axis))
-      end)
-
-    hd(names)
-  end
-
   def fft({}) do
     raise ArgumentError, "expected a tensor with rank > 0, got tensor with rank 0"
   end

torchx/lib/torchx/backend.ex

@@ -337,6 +337,16 @@ defmodule Torchx.Backend do
     |> to_nx(out)
   end
 
+  @impl true
+  def stack(out, tensors, axis) do
+    reshape = put_elem(out.shape, axis, 1)
+
+    tensors
+    |> Enum.map(&(&1 |> from_nx() |> Torchx.reshape(reshape)))
+    |> Torchx.concatenate(axis)
+    |> to_nx(out)
+  end
+
   @impl true
   def gather(out, tensor, indices, opts) do
     tensor_axes = Nx.axes(tensor)