Add sliding_window operator (#9816)

* Add windows operator

* remove TODO

* Convert ICHECKs to CHECKs

* Report errors using diagnostic context

* Use more readable CHECKs

* Remove example; move comments to test

* Revert "Remove example; move comments to test"

This is a partial revert.

This reverts commit c810c2d.

* Add newline to fix Sphinx error

* windows -> sliding_window

* whitespace

* fmt

include/tvm/relay/attrs/transform.h

@@ -33,6 +33,29 @@
 namespace tvm {
 namespace relay {
 
+/*! \brief Attributes used for the sliding_window operator */
+struct SlidingWindowAttrs : public tvm::AttrsNode<SlidingWindowAttrs> {
+  int axis;
+  Array<Integer> window_shape;
+  Array<Integer> strides;
+  TVM_DECLARE_ATTRS(SlidingWindowAttrs, "relay.attrs.SlidingWindowAttrs") {
+    TVM_ATTR_FIELD(axis).describe(
+        "What axis the sliding window begin forming over."
+        "Window will be slid over this axis and all following axes."
+        "The axis value determines the window shape (and thus, the"
+        "number of strides):"
+        "window shape and strides must both be of length"
+        "`data.ndim-axis`.");
+    TVM_ATTR_FIELD(window_shape)
+        .describe(
+            "The window shape to form over the input."
+            "Window shape must be of length `data.ndim-axis`.");
+    TVM_ATTR_FIELD(strides).describe(
+        "How to stride the window along each dimension."
+        "Strides must be of length `data.ndim-axis`.");
+  }
+};  // struct SlidingWindowAttrs
+
 /*! \brief data type cast */
 struct CastAttrs : public tvm::AttrsNode<CastAttrs> {
   DataType dtype;

include/tvm/topi/transform.h

@@ -47,6 +47,90 @@ namespace topi {
 using namespace tvm::te;
 using namespace topi::detail;
 
+/*!
+ * \brief Creates an operation to slide a window over the input x.
+ *
+ * \param x The input tensor.
+ * \param axis What axis the window begins sliding over. Window will be slid
+ * over this axis and all following axes. The axis value determines the window
+ * shape (and thus, the number of strides): window shape and strides must both
+ * be of length `data.ndim-axis`.
+ * \param window_shape The window shape to form over the input. Window shape
+ * must be of length `data.ndim-axis`.
+ * \param strides How to stride the window along each dimension. Strides must be
+ * of length `data.ndim-axis`.
+ * \param name The name of the operation
+ * \param tag The tag to mark the operation
+ *
+ * \return A Tensor whose op member is the sliding_window operation
+ */
+inline Tensor sliding_window(const Tensor& x, int axis, Array<Integer> window_shape,
+                             Array<Integer> strides, std::string name = "T_sliding_window",
+                             std::string tag = "") {
+  CHECK_GE(axis, 0);
+  auto _axis = size_t(axis);
+  CHECK_LT(_axis, x->shape.size()) << "axis must be a valid dimension index of x.";
+  CHECK_EQ(x->shape.size() - _axis, window_shape.size())
+      << "There must be a window shape for every dimension of x "
+      << "over which we are sliding the window.";
+  CHECK_EQ(strides.size(), window_shape.size()) << "Windows and strides should be the same length.";
+
+  // Compute the new shape.
+  Array<PrimExpr> new_shape;
+  // Dimensions up until `axis` remain the same.
+  for (size_t i = 0; i < _axis; ++i) {
+    new_shape.push_back(x->shape[i]);
+  }
+
+  // New dimensions which result from sliding the window in each dimension. One new dimension per
+  // window dimension.
+  for (size_t i = 0; i < window_shape.size(); ++i) {
+    // Length of the shape along this dimension.
+    auto dim_len = x->shape[_axis + i];
+    // Length of the window along this dimension.
+    auto window_len = window_shape[i];
+    // Strides along this dimension.
+    auto stride = strides[i];
+
+    new_shape.push_back(floordiv(dim_len - (window_len - 1) + stride - 1, stride));
+  }
+
+  // Dimensions comprising the window.
+  for (size_t i = 0; i < window_shape.size(); ++i) {
+    new_shape.push_back(window_shape[i]);
+  }
+
+  ICHECK(new_shape.size() == _axis + 2 * window_shape.size());
+
+  return compute(
+      new_shape,
+      [&](const Array<Var>& indices) {
+        // The index at which to index the old tensor x.
+        Array<PrimExpr> idx;
+
+        // Dimensions up until `axis` remain the same.
+        for (size_t i = 0; i < _axis; ++i) {
+          idx.push_back(indices[i]);
+        }
+
+        for (size_t i = 0; i < window_shape.size(); ++i) {
+          // Which window in this dimension we are indexing.
+          auto window_idx = indices[_axis + i];
+          // Which index within the window we are indexing.
+          auto idx_within_window = indices[_axis + window_shape.size() + i];
+          // Stride value for this dimension.
+          auto stride = strides[i];
+
+          idx.push_back(window_idx * stride + idx_within_window);
+        }
+
+        ICHECK(idx.size() == x->shape.size());
+
+        return x(idx);
+      },
+      name, tag);
+}
+
 /*!
  * \brief Creates an operation to insert new dimensions of length 1
  *

python/tvm/relay/op/_transform.py

@@ -70,6 +70,15 @@
 # concatenate
 _reg.register_schedule("concatenate", strategy.schedule_concatenate)
 
+# sliding_window
+@_reg.register_compute("sliding_window")
+def compute_sliding_window(attrs, inputs, output_type):
+    """Compute definition of sliding_window"""
+    return [topi.sliding_window(inputs[0], attrs.axis, attrs.window_shape, attrs.strides)]
+
+
+_reg.register_strategy("sliding_window", strategy.sliding_window_strategy)
+
 # strided_set
 @_reg.register_compute("strided_set")
 def compute_strided_set(attrs, inputs, output_type):

python/tvm/relay/op/strategy/generic.py

@@ -1705,6 +1705,28 @@ def uniform_strategy(attrs, inputs, out_type, target):
     return strategy
 
 
+# sliding_window
+def wrap_compute_sliding_window():
+    """Wrap sliding_window topi compute"""
+
+    def _compute_sliding_window(attrs, inputs, _):
+        return [topi.sliding_window(inputs[0], attrs.axis, attrs.window_shape, attrs.strides)]
+
+    return _compute_sliding_window
+
+
+@override_native_generic_func("sliding_window_strategy")
+def sliding_window_strategy(attrs, inputs, out_type, target):
+    """sliding_window generic strategy"""
+    strategy = _op.OpStrategy()
+    strategy.add_implementation(
+        wrap_compute_sliding_window(),
+        wrap_topi_schedule(topi.generic.schedule_extern),
+        name="sliding_window.generic",
+    )
+    return strategy
+
+
 @override_native_generic_func("normal_strategy")
 def normal_strategy(attrs, inputs, out_type, target):
     """normal generic strategy"""

python/tvm/relay/op/transform.py

@@ -25,6 +25,64 @@
 from .tensor import shape_of
 
 
+def sliding_window(data, axis, window_shape, strides):
+    """Slide a window over the data tensor.
+
+    Parameters
+    ----------
+    data : relay.Expr
+        The input data to the operator.
+
+    axis : int
+        What axis the window begins sliding over. Window will be slid over
+        this axis and all following axes. The axis value determines the window
+        shape (and thus, the number of strides): window shape and strides must
+        both be of length `data.ndim-axis`.
+
+    window_shape : List[int]
+        The window shape to form over the input. Window shape must be of length
+        `data.ndim-axis`.
+
+    strides : List[int]
+        How to stride the window along each dimension. Strides must be of length
+        `data.ndim-axis`.
+
+    Returns
+    -------
+    result : relay.Expr
+        The resulting tensor.
+
+    Examples
+    --------
+    .. code-block:: python
+
+        # Slide a window of shape (3, 4, 5) over the x tensor, beginning with
+        # dimension 1, which slides the window over the two subtensors of
+        # shape (3, 32, 32).
+        x = relay.var("x", relay.TensorType((2, 3, 32, 32), "float32"))
+        y = relay.sliding_window(x, 1, [3, 4, 5], [1, 2, 3])
+
+        data = np.random.rand(2, 3, 32, 32).astype("float32")
+        result = create_executor().evaluate(y, {x: relay.const(data)}).numpy()
+
+        # The resulting shape still has batch size 2. Each dimension in
+        # (1, 15, 10) represents the locations where we were able to
+        # form a window; that is, we were able to place the window
+        # in one place along the dimension of length 3, 15 places along
+        # the dimension of length 32 (when striding by 2), and 10 places
+        # along the second dimension of length 32 (when striding by 3).
+        # The remaining dimension (3, 4, 5) represent the formed windows.
+        assert result.shape == (2, 1, 15, 10, 3, 4, 5)
+
+        assert np.array_equal(result[0, 0, 0, 0, :, :, :], data[0, :, 0:4, 0:5])
+        assert np.array_equal(result[1, 0, 7, 3, :, :, :], data[1, :, 14:18, 9:14])
+        assert np.array_equal(result[1, 0, 14, 9, :, :, :], data[1, :, 28:32, 27:32])
+    """
+    from .. import _ffi_api as _relay_make
+
+    return _relay_make.sliding_window(data, axis, window_shape, strides)
+
+
 def cast(data, dtype):
     """Cast input tensor to data type.
 

python/tvm/topi/transform.py

@@ -971,3 +971,33 @@ def invert_permutation(data):
         r_ind = data[ind]
         result[r_ind] = ind
     return result
+
+
+def sliding_window(data, axis, window_shape, strides):
+    """Slide a window over the data tensor.
+
+    Parameters
+    ----------
+    data : relay.Expr
+        The input data to the operator.
+
+    axis : int
+        What axis the window begins sliding over. Window will be slid over
+        this axis and all following axes. The axis value determines the window
+        shape (and thus, the number of strides): window shape and strides must
+        both be of length `data.ndim-axis`.
+
+    window_shape : List[int]
+        The window shape to form over the input. Window shape must be of length
+        `data.ndim-axis`.
+
+    strides : List[int]
+        How to stride the window along each dimension. Strides must be of length
+        `data.ndim-axis`.
+
+    Returns
+    -------
+    result : relay.Expr
+        The resulting tensor.
+    """
+    return cpp.sliding_window(data, axis, window_shape, strides)

src/relay/op/tensor/transform.cc

@@ -52,6 +52,77 @@ namespace tvm {
 namespace relay {
 using tir::IntImmNode;
 
+TVM_REGISTER_NODE_TYPE(SlidingWindowAttrs);
+
+bool SlidingWindowRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
+                      const TypeReporter& reporter) {
+  // `types` contains: [data, result]
+  ICHECK_EQ(types.size(), 2);
+  const auto* data = types[0].as<TensorTypeNode>();
+  if (data == nullptr) {
+    reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+                                     << "SlidingWindow operator expects input to be of TensorType "
+                                     << "but got " << PrettyPrint(types[0]));
+    return false;
+  }
+  const auto* param = attrs.as<SlidingWindowAttrs>();
+  const int axis = param->axis;
+
+  std::vector<IndexExpr> oshape;
+
+  // Dimensions up until `axis` remain the same.
+  for (int i = 0; i < axis; ++i) {
+    oshape.emplace_back(data->shape[i]);
+  }
+
+  // New dimensions which result from sliding the window in each dimension. One new dimension per
+  // window dimension.
+  for (size_t i = 0; i < param->window_shape.size(); ++i) {
+    // Length of the shape along this dimension.
+    auto dim_len = data->shape[axis + i];
+    // Length of the window along this dimension.
+    auto window_len = param->window_shape[i];
+    // Strides along this dimension.
+    auto stride = param->strides[i];
+
+    oshape.push_back(floordiv(dim_len - (window_len - 1) + stride - 1, stride));
+  }
+
+  // Dimensions comprising the window.
+  for (size_t i = 0; i < param->window_shape.size(); ++i) {
+    oshape.push_back(param->window_shape[i]);
+  }
+
+  reporter->Assign(types[1], TensorType(oshape, data->dtype));
+  return true;
+}
+
+Array<te::Tensor> SlidingWindowCompute(const Attrs& attrs, const Array<te::Tensor>& inputs,
+                                       const Type& out_type) {
+  const SlidingWindowAttrs* param = attrs.as<SlidingWindowAttrs>();
+  ICHECK(param != nullptr);
+  return {topi::sliding_window(inputs[0], param->axis, param->window_shape, param->strides)};
+}
+
+Expr MakeSlidingWindow(Expr data, int axis, Array<Integer> window_shape, Array<Integer> strides) {
+  auto attrs = make_object<SlidingWindowAttrs>();
+  attrs->axis = axis;
+  attrs->window_shape = window_shape;
+  attrs->strides = strides;
+  static const Op& op = Op::Get("sliding_window");
+  return Call(op, {data}, Attrs(attrs), {});
+}
+
+TVM_REGISTER_GLOBAL("relay.ir.sliding_window").set_body_typed(MakeSlidingWindow);
+
+RELAY_REGISTER_OP("sliding_window")
+    .describe(R"code(Slide window over a tensor.)code" TVM_ADD_FILELINE)
+    .set_num_inputs(1)
+    .set_attrs_type<SlidingWindowAttrs>()
+    .add_argument("data", "Tensor", "The input tensor.")
+    .add_type_rel("SlidingWindow", SlidingWindowRel)
+    .set_attr<TOpPattern>("TOpPattern", kOpaque);
+
 // relay.cast
 TVM_REGISTER_NODE_TYPE(CastAttrs);
 

src/topi/transform.cc

@@ -54,6 +54,10 @@ TVM_REGISTER_GLOBAL("topi.reshape").set_body([](TVMArgs args, TVMRetValue* rv) {
   *rv = reshape(args[0], args[1]);
 });
 
+TVM_REGISTER_GLOBAL("topi.sliding_window").set_body([](TVMArgs args, TVMRetValue* rv) {
+  *rv = sliding_window(args[0], args[1], args[2], args[3]);
+});
+
 TVM_REGISTER_GLOBAL("topi.squeeze").set_body([](TVMArgs args, TVMRetValue* rv) {
   *rv = squeeze(args[0], ArrayOrInt(args[1]));
 });

tests/python/relay/test_op_level3.py

@@ -91,6 +91,32 @@ def test_cast():
     assert yy.checked_type == relay.TensorType((8, 9, 4), "int32")
 
 
+def test_sliding_window():
+    # Slide a window of shape (3, 4, 5) over the x tensor, beginning with
+    # dimension 1, which slides the window over the two subtensors of shape (3,
+    # 32, 32).
+    x = relay.var("x", relay.TensorType((2, 3, 32, 32), "float32"))
+    y = relay.sliding_window(x, 1, [3, 4, 5], [1, 2, 3])
+
+    # The resulting shape still has batch size 2. Each dimension in (1, 15, 10)
+    # represents the locations where we were able to form a window; that is, we
+    # were able to place the window in one place along the dimension of length
+    # 3, 15 places along the dimension of length 32 (when striding by 2), and 10
+    # places along the second dimension of length 32 (when striding by 3). The
+    # remaining dimensions (3, 4, 5) represent the formed windows.
+    yy = run_infer_type(y)
+    assert yy.checked_type == relay.TensorType((2, 1, 15, 10, 3, 4, 5), "float32")
+
+    data = np.random.rand(2, 3, 32, 32).astype("float32")
+    intrp = create_executor()
+    result = intrp.evaluate(y, {x: relay.const(data)})
+    result_np = result.numpy()
+    assert result_np.shape == (2, 1, 15, 10, 3, 4, 5)
+    assert np.array_equal(result_np[0, 0, 0, 0, :, :, :], data[0, :, 0:4, 0:5])
+    assert np.array_equal(result_np[1, 0, 7, 3, :, :, :], data[1, :, 14:18, 9:14])
+    assert np.array_equal(result_np[1, 0, 14, 9, :, :, :], data[1, :, 28:32, 27:32])
+
+
 def test_clip():
     a = relay.var("a", relay.TensorType((10, 4), "float32"))
     y = relay.clip(a, 1.0, 4.0)