addmm_op.cc

/* Copyright (c) 2016 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 <string>
#include <unordered_map>
#include <vector>

#include "paddle/fluid/framework/infershape_utils.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/phi/core/infermeta_utils.h"
#include "paddle/phi/infermeta/ternary.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif

namespace paddle {
namespace operators {

constexpr int kMULMKLDNNINT8 = 1;

using framework::OpKernelType;
using framework::Tensor;

class AddMMOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const {
    auto input_data_type = OperatorWithKernel::IndicateVarDataType(ctx, "X");
#ifdef PADDLE_WITH_MKLDNN
    if (this->CanMKLDNNBeUsed(ctx, input_data_type)) {
      int customized_type_value =
          framework::OpKernelType::kDefaultCustomizedTypeValue;
      if (input_data_type == framework::DataTypeTrait<int8_t>::DataType() ||
          input_data_type == framework::DataTypeTrait<uint8_t>::DataType()) {
        customized_type_value = kMULMKLDNNINT8;
      }
      return framework::OpKernelType(input_data_type,
                                     ctx.GetPlace(),
                                     framework::DataLayout::kMKLDNN,
                                     framework::LibraryType::kMKLDNN,
                                     customized_type_value);
    }
#endif
    return framework::OpKernelType(input_data_type, ctx.GetPlace());
  }
};

class AddMMOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("Input", "(Tensor), tensor to be added to the final result.");
    AddInput("X", "(Tensor), The first input tensor for mul.");
    AddInput("Y", "(Tensor), The second input tensor for mul.");
    AddOutput("Out", "(Tensor), The output tensor of addmm op.");
    AddAttr<float>("Alpha", "coefficient of x*y.").SetDefault(1.0f);
    AddAttr<float>("Beta", "coefficient of input.").SetDefault(1.0f);
    AddComment(R"DOC(
AddMM Operator.
This operator is used to perform matrix multiplication for input $x$ and $y$ with coefficient $alpha$.
$input$ with coefficient $beta$ is added to the final result.
The equation is:

$$Out = alpha * x * y + beta * input$$

$x$ and $y$ must be two-dimensional, and $input$ can be broadcastable.
)DOC");
  }
};

class AddMMGradOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE_EQ(
        ctx->HasInput("Input"),
        true,
        platform::errors::NotFound("Input(Input) should not be null"));
    PADDLE_ENFORCE_EQ(
        ctx->HasInput("X"),
        true,
        platform::errors::NotFound("Input(X) should not be null"));
    PADDLE_ENFORCE_EQ(
        ctx->HasInput("Y"),
        true,
        platform::errors::NotFound("Input(Y) should not be null"));
    PADDLE_ENFORCE_EQ(
        ctx->HasInput(framework::GradVarName("Out")),
        true,
        platform::errors::NotFound("Input(Out@GRAD) should not be null"));
    const auto& input_dims = ctx->GetInputDim("Input");
    const auto& x_dims = ctx->GetInputDim("X");
    const auto& y_dims = ctx->GetInputDim("Y");

    auto input_grad_name = framework::GradVarName("Input");
    auto x_grad_name = framework::GradVarName("X");
    auto y_grad_name = framework::GradVarName("Y");

    if (ctx->HasOutput(input_grad_name)) {
      ctx->SetOutputDim(input_grad_name, input_dims);
    }
    if (ctx->HasOutput(x_grad_name)) {
      ctx->SetOutputDim(x_grad_name, x_dims);
    }
    if (ctx->HasOutput(y_grad_name)) {
      ctx->SetOutputDim(y_grad_name, y_dims);
    }
  }
};

template <typename T>
class AddMMOpGradMaker : public framework::SingleGradOpMaker<T> {
 public:
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;

 protected:
  void Apply(GradOpPtr<T> retv) const override {
    retv->SetType("addmm_grad");
    retv->SetInput("Input", this->Input("Input"));
    retv->SetInput("X", this->Input("X"));
    retv->SetInput("Y", this->Input("Y"));
    retv->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
    retv->SetOutput(framework::GradVarName("Input"), this->InputGrad("Input"));
    retv->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    retv->SetOutput(framework::GradVarName("Y"), this->InputGrad("Y"));
    retv->SetAttrMap(this->Attrs());
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
DECLARE_INFER_SHAPE_FUNCTOR(addmm,
                            AddmmInferShapeFunctor,
                            PD_INFER_META(phi::AddmmInferMeta));
REGISTER_OPERATOR(addmm,
                  ops::AddMMOp,
                  ops::AddMMOpMaker,
                  ops::AddMMOpGradMaker<paddle::framework::OpDesc>,
                  ops::AddMMOpGradMaker<paddle::imperative::OpBase>,
                  AddmmInferShapeFunctor);

REGISTER_OPERATOR(addmm_grad, ops::AddMMGradOp);