supports collective training with programs (#18392)
1. Since allreduce op has 4 reduce types, We split these four reduce types into four ops 2. We also refined the collective op code, e.g. we separated the collective op kernel into CPUKernel and CUDAKernel, and remove the device specified DeviceContext parameter in template as we already knew the target DeviceContext 3. We remove the newly added Collective op role to reduce the complexity of program and graph analysissum_op
Yi Liu
6 years ago
committed by
GitHub
parent
85b49d8473
commit
a873fa84ce
@ -0,0 +1,39 @@
|
||||
/* Copyright (c) 2019 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/collective/c_allreduce_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class CAllReduceMaxOpMaker : public CAllReduceOpMaker {
|
||||
protected:
|
||||
std::string GetName() const override { return "Max"; }
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
namespace plat = paddle::platform;
|
||||
|
||||
REGISTER_OP_WITHOUT_GRADIENT(c_allreduce_max, ops::CAllReduceOp,
|
||||
ops::CAllReduceMaxOpMaker);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(c_allreduce_max,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMax, float>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMax, double>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMax, int>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMax, int64_t>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMax, plat::float16>);
|
||||
@ -0,0 +1,39 @@
|
||||
/* Copyright (c) 2019 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/collective/c_allreduce_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class CAllReduceMinOpMaker : public CAllReduceOpMaker {
|
||||
protected:
|
||||
std::string GetName() const override { return "Min"; }
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
namespace plat = paddle::platform;
|
||||
|
||||
REGISTER_OP_WITHOUT_GRADIENT(c_allreduce_min, ops::CAllReduceOp,
|
||||
ops::CAllReduceMinOpMaker);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(c_allreduce_min,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMin, float>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMin, double>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMin, int>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMin, int64_t>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedMin, plat::float16>);
|
||||
@ -0,0 +1,25 @@
|
||||
/* Copyright (c) 2019 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/collective/c_allreduce_op.h"
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
namespace plat = paddle::platform;
|
||||
|
||||
REGISTER_OP_CUDA_KERNEL(
|
||||
c_allreduce_min, ops::CAllReduceOpCUDAKernel<ops::kRedMin, float>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedMin, double>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedMin, int>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedMin, int64_t>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedMin, plat::float16>)
|
||||
@ -1,83 +0,0 @@
|
||||
/* Copyright (c) 2019 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 <future> // NOLINT
|
||||
#include <ostream>
|
||||
|
||||
#include "paddle/fluid/operators/collective/c_allreduce_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class CAllReduceOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||
ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
|
||||
}
|
||||
|
||||
protected:
|
||||
framework::OpKernelType GetExpectedKernelType(
|
||||
const framework::ExecutionContext& ctx) const override {
|
||||
return framework::OpKernelType(ctx.Input<framework::Tensor>("X")->type(),
|
||||
ctx.GetPlace());
|
||||
}
|
||||
};
|
||||
|
||||
class CAllReduceOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
void Make() {
|
||||
AddInput("X", "(Tensor), tensor to be allreduced.");
|
||||
AddOutput("Out", "(Tensor) the allreduced result.");
|
||||
AddAttr<int>("reduce_type", "(int default 0) determin the reduce type.")
|
||||
.SetDefault(0);
|
||||
AddAttr<int>("ring_id", "(int default 0) communication ring id.")
|
||||
.SetDefault(0);
|
||||
AddAttr<bool>(
|
||||
"use_calc_stream",
|
||||
"(bool default false) eject CUDA operations to calculation stream.")
|
||||
.SetDefault(false);
|
||||
AddComment(R"DOC(
|
||||
***CAllReduce Operator***
|
||||
|
||||
Call NCCL collective AllReduce internally. Note that this op must be used when one
|
||||
thread is managing one GPU device.
|
||||
|
||||
For speed reasons, reduce_type should be an integer:
|
||||
|
||||
0: sum
|
||||
1: prod
|
||||
2: max
|
||||
3: min
|
||||
If input and output are the same variable, in-place allreduce will be used.
|
||||
)DOC");
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
namespace plat = paddle::platform;
|
||||
|
||||
REGISTER_OP_WITHOUT_GRADIENT(c_allreduce, ops::CAllReduceOp,
|
||||
ops::CAllReduceOpMaker);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
c_allreduce, ops::CAllReduceOpKernel<plat::CPUDeviceContext, float>,
|
||||
ops::CAllReduceOpKernel<plat::CPUDeviceContext, double>,
|
||||
ops::CAllReduceOpKernel<plat::CPUDeviceContext, int>,
|
||||
ops::CAllReduceOpKernel<plat::CPUDeviceContext, int64_t>,
|
||||
ops::CAllReduceOpKernel<plat::CPUDeviceContext, plat::float16>);
|
||||
@ -0,0 +1,39 @@
|
||||
/* Copyright (c) 2019 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/collective/c_allreduce_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class CAllReduceProdOpMaker : public CAllReduceOpMaker {
|
||||
protected:
|
||||
std::string GetName() const override { return "Prod"; }
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
namespace plat = paddle::platform;
|
||||
|
||||
REGISTER_OP_WITHOUT_GRADIENT(c_allreduce_prod, ops::CAllReduceOp,
|
||||
ops::CAllReduceProdOpMaker);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(c_allreduce_prod,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedProd, float>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedProd, double>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedProd, int>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedProd, int64_t>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedProd, plat::float16>)
|
||||
@ -0,0 +1,25 @@
|
||||
/* Copyright (c) 2019 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/collective/c_allreduce_op.h"
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
namespace plat = paddle::platform;
|
||||
|
||||
REGISTER_OP_CUDA_KERNEL(
|
||||
c_allreduce_prod, ops::CAllReduceOpCUDAKernel<ops::kRedProd, float>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedProd, double>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedProd, int>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedProd, int64_t>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedProd, plat::float16>)
|
||||
@ -0,0 +1,54 @@
|
||||
/* Copyright (c) 2019 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/collective/c_allreduce_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class CAllReduceSumOpGradMaker : public framework::SingleGradOpDescMaker {
|
||||
public:
|
||||
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
|
||||
|
||||
protected:
|
||||
std::unique_ptr<framework::OpDesc> Apply() const override {
|
||||
std::unique_ptr<framework::OpDesc> retv(new framework::OpDesc());
|
||||
retv->SetType("c_allreduce_sum");
|
||||
retv->SetInput("X", OutputGrad("Out"));
|
||||
retv->SetOutput("Out", InputGrad("X"));
|
||||
retv->SetAttrMap(Attrs());
|
||||
return retv;
|
||||
}
|
||||
};
|
||||
|
||||
class CAllReduceSumOpMaker : public CAllReduceOpMaker {
|
||||
protected:
|
||||
std::string GetName() const override { return "Sum"; }
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
namespace plat = paddle::platform;
|
||||
|
||||
REGISTER_OPERATOR(c_allreduce_sum, ops::CAllReduceOp,
|
||||
ops::CAllReduceSumOpGradMaker, ops::CAllReduceSumOpMaker);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(c_allreduce_sum,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedSum, float>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedSum, double>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedSum, int>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedSum, int64_t>,
|
||||
ops::CAllReduceOpCPUKernel<ops::kRedSum, plat::float16>)
|
||||
@ -0,0 +1,25 @@
|
||||
/* Copyright (c) 2019 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/collective/c_allreduce_op.h"
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
namespace plat = paddle::platform;
|
||||
|
||||
REGISTER_OP_CUDA_KERNEL(
|
||||
c_allreduce_sum, ops::CAllReduceOpCUDAKernel<ops::kRedSum, float>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedSum, double>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedSum, int>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedSum, int64_t>,
|
||||
ops::CAllReduceOpCUDAKernel<ops::kRedSum, plat::float16>)
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue