(this PR is just for my personal self-learning) [PaddlePaddle Hackathon 2 No.22] add paddle.index_add to Paddle

paddle/fluid/operators/index_add_op.cc

@@ -0,0 +1,120 @@
+/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <memory>
+#include "paddle/fluid/framework/infershape_utils.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/op_version_registry.h"
+#include "paddle/phi/core/infermeta_utils.h"
+#include "paddle/phi/infermeta/unary.h"
+
+namespace paddle {
+namespace operators {
+
+class IndexAddOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        OperatorWithKernel::IndicateVarDataType(ctx, "X"), ctx.GetPlace());
+  }
+};
+
+class IndexAddOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor, default input Tensor<float>), "
+             "the input feature data of IndexAddOp, dtype should be"
+             "bool, int32, int64, float16, float32, float64.");
+    AddInput("IndexTensor",
+             "(Tensor, optional) If provided, index_add will use this."
+             "It has higher priority than attr(index).")
+        .AsDispensable();
+    AddInput("AxisTensor",
+             "(Tensor) If provided, use this as "
+             "axis, this has a higher priority than "
+             "attr(axis), the numel of this tensor MUST BE 1.")
+        .AsDispensable();
+    AddAttr<std::vector<int64_t>>(
+        "index",
+        "(list<int>) indices of corresponding axis in `axis`");
+    AddAttr<int>("axis", "(int), the dimension in which we index.");
+    AddAttr<float>("add_value", "(float) The value to add.");
+    AddOutput("Out",
+              "(Tensor, default Tensor<float>),"
+              " the output of  IndexAddOp, whose dtype and shape are the same as X.");
+    AddComment(R"DOC(
+                IndexAdd operator
+                Add the elements of the input tensor with value
+                by selecting the indices in the order given in index.
+                This operator also supports inplace modification.
+        )DOC");
+  }
+};
+
+template <typename T>
+class IndexAddGradMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("index_add_grad");
+    if (this->HasInput("AxisTensor")) {
+      op->SetInput("AxisTensor", this->Input("AxisTensor"));
+    }
+    if (this->HasInput("IndexTensor")) {
+      op->SetInput("IndexTensor", this->Input("IndexTensor"));
+    }
+    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+    op->SetAttrMap(this->Attrs());
+  }
+};
+
+class IndexAddGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
+                                       ctx, framework::GradVarName("Out")),
+                                   ctx.GetPlace());
+  }
+};
+
+DECLARE_INPLACE_OP_INFERER(IndexAddInplaceInferer, {"X", "Out"});
+DECLARE_INPLACE_OP_INFERER(IndexAddGradInplaceInferer,
+                           {framework::GradVarName("Out"),
+                            framework::GradVarName("X")});
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+DECLARE_INFER_SHAPE_FUNCTOR(index_add, IndexAddInferShapeFunctor,
+                            PD_INFER_META(phi::IndexAddInferMeta));
+
+REGISTER_OPERATOR(index_add, ops::IndexAddOp, ops::IndexAddOpMaker,
+                  ops::IndexAddGradMaker<paddle::framework::OpDesc>,
+                  ops::IndexAddGradMaker<paddle::imperative::OpBase>,
+                  ops::IndexAddInplaceInferer, IndexAddInferShapeFunctor);
+
+DECLARE_INFER_SHAPE_FUNCTOR(index_add_grad, IndexAddGradInferShapeFunctor,
+                            PD_INFER_META(phi::IndexAddGradInferMeta));
+
+REGISTER_OPERATOR(index_add_grad, ops::IndexAddGradOp,
+                  ops::IndexAddGradInplaceInferer,
+                  IndexAddGradInferShapeFunctor);

paddle/fluid/operators/index_add_tensor_op.cc

@@ -0,0 +1,129 @@
+/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <memory>
+#include "paddle/fluid/framework/infershape_utils.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/op_version_registry.h"
+#include "paddle/phi/core/infermeta_utils.h"
+#include "paddle/phi/infermeta/binary.h"
+
+namespace paddle {
+namespace operators {
+
+class IndexAddTensorOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        OperatorWithKernel::IndicateVarDataType(ctx, "X"), ctx.GetPlace());
+  }
+};
+
+class IndexAddTensorOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor), "
+             "the input feature data of IndexAddTensorOp, dtype should be"
+             "bool, int32, int64, float16, float32, float64.");
+    AddInput("AddValue",
+             "(Tensor>), "
+             "the input tensor of IndexAddTensorOp, dtype should be"
+             "the same as input tensor X");
+    AddInput("IndexTensor",
+             "(Tensor, optional) If provided, index_add will use this."
+             "It has higher priority than attr(index).")
+        .AsDispensable();
+    AddInput("AxisTensor",
+             "(Tensor) If provided, use this as "
+             "axis, this has a higher priority than "
+             "attr(axis), the shape of this tensor MUST BE (1,).")
+        .AsDispensable();
+    AddAttr<std::vector<int64_t>>(
+        "index",
+        "(list<int>) indices of corresponding axis in `axis`");
+    AddAttr<int>("axis", "(int), the dimension in which we index.");
+    AddOutput(
+        "Out",
+        "(Tensor, default Tensor<float>),"
+        " the output of  IndexAddTensorOp, whose dtype is the same as X.");
+    AddComment(R"DOC(
+                IndexAddTensor operator
+                Add the elements of the input tensor with value
+                by selecting the indices in the order given in index.
+                This operator also supports inplace modification.
+        )DOC");
+  }
+};
+
+template <typename T>
+class IndexAddTensorGradMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("index_add_tensor_grad");
+    if (this->HasInput("AxisTensor")) {
+      op->SetInput("AxisTensor", this->Input("AxisTensor"));
+    }
+    if (this->HasInput("IndexTensor")) {
+      op->SetInput("IndexTensor", this->Input("IndexTensor"));
+    }
+    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+    op->SetOutput(framework::GradVarName("AddValue"),
+                  this->InputGrad("AddValue"));
+    op->SetAttrMap(this->Attrs());
+  }
+};
+
+class IndexAddTensorGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
+                                       ctx, framework::GradVarName("Out")),
+                                   ctx.GetPlace());
+  }
+};
+
+DECLARE_INPLACE_OP_INFERER(IndexAddTensorInplaceInferer, {"X", "Out"});
+DECLARE_INPLACE_OP_INFERER(IndexAddTensorGradInplaceInferer,
+                           {framework::GradVarName("Out"),
+                            framework::GradVarName("X")});
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+DECLARE_INFER_SHAPE_FUNCTOR(index_add_tensor, IndexAddTensorInferShapeFunctor,
+                            PD_INFER_META(phi::IndexAddTensorInferMeta));
+
+REGISTER_OPERATOR(index_add_tensor, ops::IndexAddTensorOp,
+                  ops::IndexAddTensorOpMaker,
+                  ops::IndexAddTensorGradMaker<paddle::framework::OpDesc>,
+                  ops::IndexAddTensorGradMaker<paddle::imperative::OpBase>,
+                  ops::IndexAddTensorInplaceInferer,
+                  IndexAddTensorInferShapeFunctor);
+
+DECLARE_INFER_SHAPE_FUNCTOR(index_add_tensor_grad,
+                            IndexAddTensorGradInferShapeFunctor,
+                            PD_INFER_META(phi::IndexAddTensorGradInferMeta));
+
+REGISTER_OPERATOR(index_add_tensor_grad, ops::IndexAddTensorGradOp,
+                  ops::IndexAddTensorGradInplaceInferer,
+                  IndexAddTensorGradInferShapeFunctor);

paddle/phi/infermeta/binary.cc

@@ -1181,6 +1181,62 @@ void HuberLossInferMeta(const MetaTensor& input,
   out->share_lod(input);
 }
 
+void IndexAddTensorInferMeta(const MetaTensor& x,
+                              const MetaTensor& add_tensor,
+                              const IntArray& index_arr,
+                              const Scalar& axis_scalar,
+                              MetaTensor* output) {
+  auto input_dim = x.dims();
+  auto index = index_arr.GetData();
+  auto axis = axis_scalar.to<int>();
+
+  PADDLE_ENFORCE_EQ(
+      add_tensor.dtype() == x.dtype(),
+      true,
+      phi::errors::InvalidArgument(
+          "The dtype of add_tensor should be same as input tensor."));
+
+  PADDLE_ENFORCE_EQ(
+      axis < input_dim.size() && axis >= (0 - input_dim.size()),
+      true,
+      phi::errors::OutOfRange(
+          "Axis is out of range, It's expected "
+          "to be in range of [-%d, %d). But received Attr(axis) = %d.",
+          input_dim.size(),
+          input_dim.size(),
+          axis));
+
+  PADDLE_ENFORCE_EQ(
+      index.size() > 0,
+      true,
+      phi::errors::InvalidArgument("The index array should not be empty."));
+
+  output->set_dims(x.dims());
+  output->set_dtype(x.dtype());
+  output->set_layout(x.layout());
+  output->share_lod(x);
+}
+
+void IndexAddTensorGradInferMeta(const MetaTensor& out_grad,
+                                  const IntArray& index_arr,
+                                  const Scalar& axis_scalar,
+                                  MetaTensor* x_grad,
+                                  MetaTensor* add_tensor_grad) {
+  auto do_dims = out_grad.dims();
+  if (x_grad) {
+    x_grad->set_dims(do_dims);
+    x_grad->set_dtype(out_grad.dtype());
+    x_grad->set_layout(out_grad.layout());
+    x_grad->share_lod(out_grad);
+  }
+
+  if (add_tensor_grad) {
+    add_tensor_grad->set_dims(phi::make_ddim({1}));
+    add_tensor_grad->set_dtype(out_grad.dtype());
+  }
+}
+
+
 void IndexSampleInferMeta(const MetaTensor& x,
                           const MetaTensor& y,
                           MetaTensor* out,
@@ -1534,7 +1590,7 @@ void MvInferMeta(const MetaTensor& x, const MetaTensor& vec, MetaTensor* out) {
                     phi::errors::InvalidArgument(
                         "X's second dimension is expected to be equal to "
                         "Vec's first dimension"
-                        "but received X'shape = [%s], Vec's shape = [%s]",
+                        "but recieved X'shape = [%s], Vec's shape = [%s]",
                         dim_x,
                         dim_vec));
 

paddle/phi/infermeta/binary.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/phi/common/int_array.h"
 #include "paddle/phi/common/scalar.h"
 #include "paddle/phi/core/meta_tensor.h"
 
@@ -184,6 +185,18 @@ void HuberLossInferMeta(const MetaTensor& input_meta,
                         MetaTensor* residual,
                         MetaConfig config = MetaConfig());
 
+void IndexAddTensorInferMeta(const MetaTensor& x,
+                              const MetaTensor& add_tensor,
+                              const IntArray& index_arr,
+                              const Scalar& axis_scalar,
+                              MetaTensor* output);
+
+void IndexAddTensorGradInferMeta(const MetaTensor& out_grad,
+                                  const IntArray& index_arr,
+                                  const Scalar& axis_scalar,
+                                  MetaTensor* x_grad,
+                                  MetaTensor* add_tensor_grad);
+
 void IndexSampleInferMeta(const MetaTensor& x,
                           const MetaTensor& y,
                           MetaTensor* out,