tile and repeat operator added in rely

Author: Laurawly

docs/api/python/topi.rst

@@ -72,6 +72,8 @@ List of operators
    topi.logical_or
    topi.logical_not
    topi.arange
+   topi.repeat
+   topi.tile
    topi.layout_transform
    topi.image.resize
 
@@ -130,6 +132,8 @@ topi
 .. autofunction:: topi.greater
 .. autofunction:: topi.less
 .. autofunction:: topi.arange
+.. autofunction:: topi.repeat
+.. autofunction:: topi.tile
 .. autofunction:: topi.layout_transform
 
 topi.nn

docs/langref/relay_op.rst

@@ -96,6 +96,8 @@ This level enables additional math and transform operators.
    tvm.relay.cast
    tvm.relay.split
    tvm.relay.arange
+   tvm.relay.repeat
+   tvm.relay.tile
 
 
 **Level 4: Broadcast and Reductions**
@@ -220,6 +222,8 @@ Level 3 Definitions
 .. autofunction:: tvm.relay.cast
 .. autofunction:: tvm.relay.split
 .. autofunction:: tvm.relay.arange
+.. autofunction:: tvm.relay.repeat
+.. autofunction:: tvm.relay.tile
 
 
 Level 4 Definitions

include/tvm/relay/attrs/transform.h

@@ -115,6 +115,28 @@ struct ArangeAttrs : public tvm::AttrsNode<ArangeAttrs> {
   }
 };  // struct ArangeAttrs
 
+/*! \brief Attributes used in repeat operators */
+struct RepeatAttrs : public tvm::AttrsNode<RepeatAttrs> {
+  Integer repeats;
+  Integer axis;
+  TVM_DECLARE_ATTRS(RepeatAttrs, "relay.attrs.RepeatAttrs") {
+    TVM_ATTR_FIELD(repeats)
+        .describe("The number of repetitions for each element.");
+    TVM_ATTR_FIELD(axis).set_default(NullValue<Integer>())
+        .describe(" The axis along which to repeat values.");
+  }
+};  // struct RepeatAttrs
+
+/*! \brief Attributes used in tile operators */
+struct TileAttrs : public tvm::AttrsNode<TileAttrs> {
+  Array<Integer> reps;
+  TVM_DECLARE_ATTRS(TileAttrs, "relay.attrs.TileAttrs") {
+    TVM_ATTR_FIELD(reps)
+        .describe("The number of times for repeating the tensor a."
+                  "Each dim sizeof reps must be a positive integer.");
+  }
+};  // struct TileAttrs
+
 /*! \brief Attributes used in squeeze operators */
 struct SqueezeAttrs : public tvm::AttrsNode<SqueezeAttrs> {
   // use axis to make the name numpy compatible.

python/tvm/relay/frontend/mxnet.py

@@ -166,6 +166,10 @@ def _mx_dropout(inputs, attrs):
     return _op.nn.dropout(inputs[0], rate=rate)
 
 
+def _mx_BlockGrad(inputs, attrs):
+    return inputs
+
+
 def _mx_batch_norm(inputs, attrs):
     if attrs.get_bool("output_mean_var", False):
         raise RuntimeError("batch_norm do not support output_mean_var")
@@ -314,6 +318,21 @@ def _mx_arange(inputs, attrs):
     return _op.arange(**new_attrs)
 
 
+def _mx_repeat(inputs, attrs):
+    assert len(inputs) == 1
+    new_attrs = {}
+    new_attrs["repeats"] = attrs.get_int("repeats")
+    new_attrs["axis"] = attrs.get_int("axis", 0)
+    return _op.repeat(inputs[0], **new_attrs)
+
+
+def _mx_tile(inputs, attrs):
+    assert len(inputs) == 1
+    new_attrs = {}
+    new_attrs["reps"] = attrs.get_int_tuple("reps")
+    return _op.tile(inputs[0], **new_attrs)
+
+
 def _mx_roi_align(inputs, attrs):
     new_attrs = {}
     new_attrs["pooled_size"] = attrs.get_int_tuple("pooled_size")
@@ -431,6 +450,9 @@ def _mx_roi_align(inputs, attrs):
     "batch_dot"     : _mx_batch_dot,
     "LeakyReLU"     : _mx_leaky_relu,
     "_arange"       : _mx_arange,
+    "repeat"        : _mx_repeat,
+    "tile"          : _mx_tile,
+    "BlockGrad"     : _mx_BlockGrad,
     "SoftmaxOutput" : _mx_softmax_output,
     "SoftmaxActivation" : _mx_softmax_activation,
     # vision

python/tvm/relay/op/_transform.py

@@ -20,6 +20,8 @@
 _reg.register_schedule("full", schedule_injective)
 _reg.register_schedule("full_like", schedule_injective)
 _reg.register_schedule("arange", schedule_injective)
+_reg.register_schedule("repeat", schedule_broadcast)
+_reg.register_schedule("tile", schedule_broadcast)
 _reg.register_schedule("cast", schedule_injective)
 _reg.register_schedule("strided_slice", schedule_injective)
 _reg.register_schedule("slice_like", schedule_injective)

python/tvm/relay/op/transform.py

@@ -294,6 +294,56 @@ def arange(start, stop=None, step=1, dtype="float32"):
     return _make.arange(start, stop, step, dtype)
 
 
+def repeat(data, repeats, axis):
+    """Repeats elements of an array.
+    By default, repeat flattens the input array into 1-D and then repeats the elements.
+
+    Parameters
+    ----------
+    data : relay.Expr
+        The input data to the operator.
+
+    repeats : int
+        The number of repetitions for each element.
+
+    axis: int
+        The axis along which to repeat values. The negative numbers are interpreted
+        counting from the backward. By default, use the flattened input array, and
+        return a flat output array.
+
+    Returns
+    -------
+    ret : relay.Expr
+        The computed result.
+    """
+    return _make.repeat(data, repeats, axis)
+
+
+def tile(data, reps):
+    """Repeats the whole array multiple times.
+
+    Parameters
+    ----------
+    data : relay.Expr
+        The input data to the operator.
+
+    reps : tuple of int
+        The number of times repeating the tensor a.
+
+    .. note::
+        Each dim size of reps must be a positive integer. If reps has length d,
+        the result will have dimension of max(d, a.ndim); If a.ndim < d, a is
+        promoted to be d-dimensional by prepending new axes. If a.ndim ? d, reps
+        is promoted to a.ndim by pre-pending 1's to it.
+
+    Returns
+    -------
+    ret : relay.Expr
+        The computed result.
+    """
+    return _make.tile(data, reps)
+
+
 def where(condition, x, y):
     """Selecting elements from either x or y depending on the value of the
     condition.

src/relay/op/tensor/transform.cc

@@ -937,6 +937,175 @@ RELAY_REGISTER_OP("arange")
 .set_attr<FTVMCompute>("FTVMCompute", ArangeCompute)
 .set_attr<TOpPattern>("TOpPattern", kInjective);
 
+// repeat operator
+TVM_REGISTER_NODE_TYPE(RepeatAttrs);
+
+bool RepeatRel(const Array<Type>& types,
+                   int num_inputs,
+                   const Attrs& attrs,
+                   const TypeReporter& reporter) {
+  // `types` contains: [data, result]
+  CHECK_EQ(types.size(), 2);
+  const auto* data = types[0].as<TensorTypeNode>();
+  if (data == nullptr) {
+    CHECK(types[0].as<IncompleteTypeNode>())
+        << "repeat: expect input type to be TensorType but get "
+        << types[0];
+    return false;
+  }
+  const auto* param = attrs.as<RepeatAttrs>();
+  const int ndim = static_cast<int>(data->shape.size());
+  const int repeats = param->repeats;
+  const int axis = param->axis;
+  CHECK(repeats >= 1)
+    << "repeat only accepts `repeats >= 1`"
+    << ", but got repeats = " << repeats;
+  CHECK(-ndim - 1 <= axis && axis <= ndim)
+    << "repeat only accepts `axis` in [-data.ndim - 1, data.ndim]"
+    << ", but got axis = " << axis
+    << ", and data.ndim = " << ndim;
+  const int pivot = axis < 0 ? ndim + axis : axis;
+  std::vector<IndexExpr> oshape;
+  oshape.reserve(ndim + repeats);
+  for (int i = 0; i < pivot; ++i) {
+    oshape.emplace_back(data->shape[i]);
+  }
+  oshape.emplace_back(data->shape[pivot] * repeats);
+  for (int i = pivot + 1; i < ndim; ++i) {
+    oshape.emplace_back(data->shape[i]);
+  }
+  reporter->Assign(types[1], TensorTypeNode::make(oshape, data->dtype));
+  return true;
+}
+
+Array<Tensor> RepeatCompute(const Attrs& attrs,
+                            const Array<Tensor>& inputs,
+                            const Type& out_type,
+                            const Target& target) {
+  const RepeatAttrs *param = attrs.as<RepeatAttrs>();
+  CHECK(param != nullptr);
+  return { topi::repeat(inputs[0], param->repeats, param->axis) };
+}
+
+Expr MakeRepeat(Expr data,
+                    int repeats,
+                    int axis) {
+  auto attrs = make_node<RepeatAttrs>();
+  attrs->repeats = repeats;
+  attrs->axis = axis;
+  static const Op& op = Op::Get("repeat");
+  return CallNode::make(op, {data}, Attrs(attrs), {});
+}
+
+TVM_REGISTER_API("relay.op._make.repeat")
+.set_body([](const TVMArgs& args, TVMRetValue* rv) {
+    runtime::detail::unpack_call<Expr, 3>(MakeRepeat, args, rv);
+});
+
+RELAY_REGISTER_OP("repeat")
+.describe(R"code(Repeat elements of an array `repeats` times along axis `axis`
+
+- **data**: The input data to the operator.
+
+)code" TVM_ADD_FILELINE)
+.set_num_inputs(1)
+.set_attrs_type_key("relay.attrs.Repeat")
+.add_argument("data", "Tensor", "The input tensor.")
+.set_support_level(1)
+.add_type_rel("Repeat", RepeatRel)
+.set_attr<FTVMCompute>("FTVMCompute", RepeatCompute)
+.set_attr<TOpPattern>("TOpPattern", kBroadcast);
+
+// tile operator
+TVM_REGISTER_NODE_TYPE(TileAttrs);
+
+bool TileRel(const Array<Type>& types,
+             int num_inputs,
+             const Attrs& attrs,
+             const TypeReporter& reporter) {
+  // `types` contains: [data, result]
+  CHECK_EQ(types.size(), 2);
+  const auto* data = types[0].as<TensorTypeNode>();
+  if (data == nullptr) {
+    CHECK(types[0].as<IncompleteTypeNode>())
+        << "tile: expect input type to be TensorType but get "
+        << types[0];
+    return false;
+  }
+  const auto* param = attrs.as<TileAttrs>();
+  const int ndim = static_cast<int>(data->shape.size());
+  const Array<Integer>& reps = param->reps;
+  // check dimension match
+  CHECK(!reps.defined())
+    << "repetition array is not defined. data.ndim = " << ndim;
+  const int rndim = static_cast<int>(reps.size());
+  int tndim = (ndim > rndim) ? ndim : rndim;
+  // re-construct data shape or reps shape
+  std::vector<IndexExpr> data_shape;
+  std::vector<IndexExpr> reps_shape;
+  data_shape.reserve(tndim);
+  reps_shape.reserve(tndim);
+  if (ndim == rndim) {
+    for (int i = 0; i < tndim; ++i) {
+        data_shape.emplace_back(data->shape[i]);
+        reps_shape.emplace_back(reps[i]);
+    }
+  } else if (ndim > rndim) {
+    for (int i = 0; i < ndim; ++i)
+        data_shape.emplace_back(data->shape[i]);
+    for (int i = 0; i < (ndim - rndim); ++i)
+        reps_shape.emplace_back(1);
+    for (int i = 0; i < rndim; ++i)
+        reps_shape.emplace_back(reps[i]);
+  } else {
+    for (int i = 0; i < rndim; ++i)
+        reps_shape.emplace_back(reps[i]);
+  }
+  std::vector<IndexExpr> oshape;
+  oshape.reserve(tndim);
+  for (int i = 0; i < tndim; ++i) {
+    oshape.emplace_back(data_shape[i] * reps_shape[i]);
+  }
+  reporter->Assign(types[1], TensorTypeNode::make(oshape, data->dtype));
+  return true;
+}
+
+Array<Tensor> TileCompute(const Attrs& attrs,
+                          const Array<Tensor>& inputs,
+                          const Type& out_type,
+                          const Target& target) {
+  const TileAttrs *param = attrs.as<TileAttrs>();
+  CHECK(param != nullptr);
+  return { topi::tile(inputs[0], param->reps) };
+}
+
+Expr MakeTile(Expr data,
+              Array<Integer> reps) {
+  auto attrs = make_node<TileAttrs>();
+  attrs->reps = reps;
+  static const Op& op = Op::Get("tile");
+  return CallNode::make(op, {data}, Attrs(attrs), {});
+}
+
+TVM_REGISTER_API("relay.op._make.tile")
+.set_body([](const TVMArgs& args, TVMRetValue* rv) {
+    runtime::detail::unpack_call<Expr, 2>(MakeTile, args, rv);
+});
+
+RELAY_REGISTER_OP("tile")
+.describe(R"code(Repeat the whole array multiple times.
+
+- **data**: The input data to the operator.
+
+)code" TVM_ADD_FILELINE)
+.set_num_inputs(1)
+.set_attrs_type_key("relay.attrs.Tile")
+.add_argument("data", "Tensor", "The input tensor.")
+.set_support_level(1)
+.add_type_rel("Tile", TileRel)
+.set_attr<FTVMCompute>("FTVMCompute", TileCompute)
+.set_attr<TOpPattern>("TOpPattern", kBroadcast);
+
 // where operator
 bool WhereRel(const Array<Type>& types,
               int num_inputs,