提交 7b6bf9dd 编写于 作者: X Xin Pan

make fill_constant kernel-based

test=develop
上级 61491ce2
......@@ -12,103 +12,40 @@ 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 "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/fill_constant_op.h"
namespace paddle {
namespace operators {
class FillConstantInferShape : public framework::InferShapeBase {
class FillConstantOp : public framework::OperatorWithKernel {
public:
void operator()(framework::InferShapeContext *ctx) const override {
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of FillConstantOp should not be null.");
auto &shape = ctx->Attrs().Get<std::vector<int64_t>>("shape");
auto& shape = ctx->Attrs().Get<std::vector<int64_t>>("shape");
ctx->SetOutputDim("Out", framework::make_ddim(shape));
}
};
class FillConstantOp : public framework::OperatorBase {
public:
using framework::OperatorBase::OperatorBase;
private:
void RunImpl(const framework::Scope &scope,
const platform::Place &dev_place) const override {
auto data_type =
static_cast<framework::proto::VarType::Type>(Attr<int>("dtype"));
auto value = Attr<float>("value");
auto force_cpu = Attr<bool>("force_cpu");
framework::Tensor *tensor = nullptr;
auto &out_var = *scope.FindVar(Output("Out"));
if (out_var.IsType<framework::LoDTensor>()) {
tensor = out_var.GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
} else if (out_var.IsType<framework::SelectedRows>()) {
tensor = out_var.GetMutable<framework::SelectedRows>()->mutable_value();
tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
} else {
PADDLE_THROW(
"fill constant op's output only"
"supports SelectedRows and LoDTensor");
}
if (force_cpu) {
auto cpu = platform::CPUPlace();
tensor->mutable_data(cpu, data_type);
} else {
tensor->mutable_data(dev_place, data_type);
}
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
auto &dev_ctx = *pool.Get(dev_place);
math::set_constant(dev_ctx, tensor, value);
}
void RunImplPrepared(const framework::RuntimeContext &ctx,
const platform::Place &dev_place) const override {
auto data_type =
static_cast<framework::proto::VarType::Type>(Attr<int>("dtype"));
auto value = Attr<float>("value");
auto force_cpu = Attr<bool>("force_cpu");
framework::Tensor *tensor = nullptr;
auto &out_var = *ctx.outputs.at("Out")[0];
if (out_var.IsType<framework::LoDTensor>()) {
tensor = out_var.GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
} else if (out_var.IsType<framework::SelectedRows>()) {
tensor = out_var.GetMutable<framework::SelectedRows>()->mutable_value();
tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
} else {
PADDLE_THROW(
"fill constant op's output only"
"supports SelectedRows and LoDTensor");
}
if (force_cpu) {
auto cpu = platform::CPUPlace();
tensor->mutable_data(cpu, data_type);
} else {
tensor->mutable_data(dev_place, data_type);
}
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
auto &dev_ctx = *pool.Get(dev_place);
math::set_constant(dev_ctx, tensor, value);
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::proto::VarType::Type(ctx.Attr<int>("dtype")),
ctx.GetPlace());
}
};
class FillConstantOpVarTypeInference : public framework::VarTypeInference {
public:
void operator()(const framework::OpDesc &op_desc,
framework::BlockDesc *block) const override {}
void operator()(const framework::OpDesc& op_desc,
framework::BlockDesc* block) const override {
auto data_type = static_cast<framework::proto::VarType::Type>(
boost::get<int>(op_desc.GetAttr("dtype")));
auto& out_var_name = op_desc.Output("Out").front();
block->Var(out_var_name)->SetDataType(data_type);
}
};
class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
......@@ -142,7 +79,11 @@ Fill up a variable with specified constant value.
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(fill_constant, ops::FillConstantOp,
ops::FillConstantInferShape, ops::FillConstantOpMaker,
paddle::framework::EmptyGradOpMaker,
ops::FillConstantOpVarTypeInference);
REGISTER_OPERATOR(fill_constant, ops::FillConstantOp, ops::FillConstantOpMaker,
ops::FillConstantOpVarTypeInference,
paddle::framework::EmptyGradOpMaker);
REGISTER_OP_CPU_KERNEL(fill_constant, ops::FillConstantKernel<float>,
ops::FillConstantKernel<double>,
ops::FillConstantKernel<int64_t>);
/* Copyright (c) 2018 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 "paddle/fluid/operators/fill_constant_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(fill_constant, ops::FillConstantKernel<float>,
ops::FillConstantKernel<double>,
ops::FillConstantKernel<int64_t>);
/* Copyright (c) 2018 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. */
#pragma once
#include <vector>
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace operators {
template <typename T>
class FillConstantKernel : public framework::OpKernel<T> {
public:
void Compute(const paddle::framework::ExecutionContext &ctx) const override {
auto data_type =
static_cast<framework::proto::VarType::Type>(ctx.Attr<int>("dtype"));
auto value = ctx.Attr<float>("value");
auto force_cpu = ctx.Attr<bool>("force_cpu");
framework::Tensor *tensor = nullptr;
framework::Variable *out_var = ctx.OutputVar("Out");
if (out_var->IsType<framework::LoDTensor>()) {
tensor = out_var->GetMutable<framework::LoDTensor>();
tensor->Resize(
framework::make_ddim(ctx.Attr<std::vector<int64_t>>("shape")));
} else if (out_var->IsType<framework::SelectedRows>()) {
tensor = out_var->GetMutable<framework::SelectedRows>()->mutable_value();
tensor->Resize(
framework::make_ddim(ctx.Attr<std::vector<int64_t>>("shape")));
} else {
PADDLE_THROW(
"fill constant op's output only"
"supports SelectedRows and LoDTensor");
}
if (force_cpu) {
tensor->mutable_data(platform::CPUPlace(), data_type);
} else {
tensor->mutable_data(ctx.GetPlace(), data_type);
}
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
auto &dev_ctx = *pool.Get(ctx.GetPlace());
math::set_constant(dev_ctx, tensor, value);
}
};
} // namespace operators
} // namespace paddle
......@@ -14,7 +14,6 @@ limitations under the License. */
#include "paddle/fluid/pybind/imperative.h"
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/imperative/tracer.h"
namespace paddle {
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册