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 analysis

paddle/fluid/framework/op_proto_maker.cc

@@ -74,7 +74,6 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
                static_cast<int>(OpRole::kBackward),
            static_cast<int>(OpRole::kOptimize) |
                static_cast<int>(OpRole::kLRSched),
-           static_cast<int>(OpRole::kCollective),
            static_cast<int>(OpRole::kNotSpecified)})
       .SetDefault(static_cast<int>(OpRole::kNotSpecified));
   AddAttr<std::vector<std::string>>(OpRoleVarAttrName(),

paddle/fluid/framework/op_proto_maker.h

@@ -34,9 +34,6 @@ enum class OpRole {
   kDist = 0x0008,
   // Tag all learning rate scheduler operators.
   kLRSched = 0x0010,
-  // Collective role is for all collective operators and other operators used
-  // for collective training
-  kCollective = 0x0020,
 
   kLoss = 0x0100,
   // The default value of op's role. This should be only used for unittests and

paddle/fluid/operators/collective/c_allgather_op.cc

@@ -13,9 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/collective/c_allgather_op.h"
-#include <future>  // NOLINT
+
 #include <memory>
-#include <ostream>
 
 namespace paddle {
 namespace operators {
@@ -25,8 +24,7 @@ class CAllGatherOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
   void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
-    PADDLE_ENFORCE(ctx->HasOutput("Out"),
-                   "Output(Out) of SyncFCGather op should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null.");
     int nranks = ctx->Attrs().Get<int>("nranks");
     PADDLE_ENFORCE_GE(nranks, 2, "nranks should be >=2");
     framework::DDim dim = ctx->GetInputDim("X");
@@ -49,10 +47,10 @@ class CAllGatherOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<int>("nranks",
                  "Total trainer count of the distributed training job");
     AddComment(R"DOC(
-***CAllGather Operator***
+CAllGather Operator
 each rank receives the aggregation of data from all ranks in the order of the ranks
 
-Call NCCL collective AllGather internally.https://docs.nvidia.com/deeplearning/sdk/nccl-developer-guide/docs/api/colls.html#c.ncclAllGather
+reference: https://docs.nvidia.com/deeplearning/sdk/nccl-developer-guide/docs/usage/operations.html#allgather
 )DOC");
   }
 };
@@ -81,9 +79,8 @@ namespace plat = paddle::platform;
 REGISTER_OPERATOR(c_allgather, ops::CAllGatherOp, ops::CAllGatherOpGradMaker,
                   ops::CAllGatherOpMaker);
 
-REGISTER_OP_CPU_KERNEL(
-    c_allgather, ops::CAllGatherOpKernel<plat::CPUDeviceContext, float>,
-    ops::CAllGatherOpKernel<plat::CPUDeviceContext, double>,
-    ops::CAllGatherOpKernel<plat::CPUDeviceContext, int>,
-    ops::CAllGatherOpKernel<plat::CPUDeviceContext, int64_t>,
-    ops::CAllGatherOpKernel<plat::CPUDeviceContext, plat::float16>);
+REGISTER_OP_CPU_KERNEL(c_allgather, ops::CAllGatherOpCPUKernel<float>,
+                       ops::CAllGatherOpCPUKernel<double>,
+                       ops::CAllGatherOpCPUKernel<int>,
+                       ops::CAllGatherOpCPUKernel<int64_t>,
+                       ops::CAllGatherOpCPUKernel<plat::float16>);

paddle/fluid/operators/collective/c_allgather_op.cu.cc

@@ -14,12 +14,64 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/collective/c_allgather_op.h"
 
+#include <memory>
+
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#include "paddle/fluid/platform/collective_helper.h"
+#include "paddle/fluid/platform/nccl_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class CAllGatherOpCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+    auto in = ctx.Input<framework::Tensor>("X");
+    auto out = ctx.Output<framework::Tensor>("Out");
+    ncclDataType_t dtype = platform::ToNCCLDataType(in->type());
+
+    int nranks = ctx.Attr<int>("nranks");
+    int rid = ctx.Attr<int>("ring_id");
+    auto comm = platform::NCCLCommContext::Instance().Get(rid);
+    PADDLE_ENFORCE_EQ(nranks, comm->nranks());
+
+    auto place = ctx.GetPlace();
+    framework::DDim out_dims = in->dims();
+    out_dims[0] *= nranks;
+    out->mutable_data<T>(out_dims, place);
+
+    int64_t send_numel = in->numel();
+    const T* send_buff = in->data<T>();
+    T* recv_buff = out->data<T>();
+
+    cudaStream_t stream = nullptr;
+    if (ctx.Attr<bool>("use_calc_stream")) {
+      auto dev_ctx = platform::DeviceContextPool::Instance().Get(place);
+      stream = static_cast<platform::CUDADeviceContext*>(dev_ctx)->stream();
+    } else {
+      stream = comm->stream();
+    }
+
+    PADDLE_ENFORCE(platform::dynload::ncclAllGather(
+        send_buff, recv_buff, send_numel, static_cast<ncclDataType_t>(dtype),
+        comm->comm(), stream));
+#else
+    PADDLE_THROW("PaddlePaddle should compile with GPU.");
+#endif
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
-REGISTER_OP_CUDA_KERNEL(
-    c_allgather, ops::CAllGatherOpKernel<plat::CUDADeviceContext, float>,
-    ops::CAllGatherOpKernel<plat::CUDADeviceContext, double>,
-    ops::CAllGatherOpKernel<plat::CUDADeviceContext, int>,
-    ops::CAllGatherOpKernel<plat::CUDADeviceContext, int64_t>,
-    ops::CAllGatherOpKernel<plat::CUDADeviceContext, plat::float16>);
+REGISTER_OP_CUDA_KERNEL(c_allgather, ops::CAllGatherOpCUDAKernel<float>,
+                        ops::CAllGatherOpCUDAKernel<double>,
+                        ops::CAllGatherOpCUDAKernel<int>,
+                        ops::CAllGatherOpCUDAKernel<int64_t>,
+                        ops::CAllGatherOpCUDAKernel<plat::float16>);

paddle/fluid/operators/collective/c_allgather_op.h

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+/* 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.
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+
 #include <algorithm>
 #include <utility>
 #include <vector>
@@ -22,52 +23,14 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-#include "paddle/fluid/platform/collective_helper.h"
-#include "paddle/fluid/platform/nccl_helper.h"
-#endif
-
 namespace paddle {
 namespace operators {
 
-template <typename DeviceContext, typename T>
-class CAllGatherOpKernel : public framework::OpKernel<T> {
+template <typename T>
+class CAllGatherOpCPUKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto place = ctx.GetPlace();
-    PADDLE_ENFORCE(is_gpu_place(place),
-                   "CAllGatherOp can run on gpu place only for now.");
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-    auto in = ctx.Input<framework::Tensor>("X");
-    auto out = ctx.Output<framework::Tensor>("Out");
-    ncclDataType_t dtype = platform::ToNCCLDataType(in->type());
-
-    int rid = ctx.Attr<int>("ring_id");
-    auto comm = platform::NCCLCommContext::Instance().Get(rid);
-    int nranks = comm->nranks();
-
-    framework::DDim out_dims = in->dims();
-    out_dims[0] *= nranks;
-    out->mutable_data<T>(out_dims, place);
-
-    int64_t send_numel = in->numel();
-    const T* send_buff = in->data<T>();
-    T* recv_buff = out->data<T>();
-
-    cudaStream_t stream = nullptr;
-    if (ctx.Attr<bool>("use_calc_stream")) {
-      auto dev_ctx = platform::DeviceContextPool::Instance().Get(place);
-      stream = static_cast<platform::CUDADeviceContext*>(dev_ctx)->stream();
-    } else {
-      stream = comm->stream();
-    }
-
-    PADDLE_ENFORCE(platform::dynload::ncclAllGather(
-        send_buff, recv_buff, send_numel, static_cast<ncclDataType_t>(dtype),
-        comm->comm(), stream));
-#else
-    PADDLE_THROW("PaddlePaddle should compile with GPU.");
-#endif
+    PADDLE_THROW("unimplemented cpu kernel for CAllGatherOp.");
   }
 };
 

paddle/fluid/operators/collective/c_allreduce_max_op.cc

+/* 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>);

paddle/fluid/operators/collective/c_allreduce_op.cu.cc → paddle/fluid/operators/collective/c_allreduce_max_op.cu.cc

@@ -18,8 +18,8 @@ namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
 REGISTER_OP_CUDA_KERNEL(
-    c_allreduce, ops::CAllReduceOpKernel<plat::CUDADeviceContext, float>,
-    ops::CAllReduceOpKernel<plat::CUDADeviceContext, double>,
-    ops::CAllReduceOpKernel<plat::CUDADeviceContext, int>,
-    ops::CAllReduceOpKernel<plat::CUDADeviceContext, int64_t>,
-    ops::CAllReduceOpKernel<plat::CUDADeviceContext, plat::float16>);
+    c_allreduce_max, ops::CAllReduceOpCUDAKernel<ops::kRedMax, float>,
+    ops::CAllReduceOpCUDAKernel<ops::kRedMax, double>,
+    ops::CAllReduceOpCUDAKernel<ops::kRedMax, int>,
+    ops::CAllReduceOpCUDAKernel<ops::kRedMax, int64_t>,
+    ops::CAllReduceOpCUDAKernel<ops::kRedMax, plat::float16>)

paddle/fluid/operators/collective/c_allreduce_min_op.cc

+/* 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>);

paddle/fluid/operators/collective/c_allreduce_min_op.cu.cc

+/* 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>)

paddle/fluid/operators/collective/c_allreduce_op.h

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+/* 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.
@@ -13,9 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include <algorithm>
-#include <utility>
-#include <vector>
+
+#include <string>
 
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/lod_tensor.h"
@@ -29,17 +28,41 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename DeviceContext, typename T>
-class CAllReduceOpKernel : public framework::OpKernel<T> {
+enum ReduceType { kRedSum, kRedMax, kRedMin, kRedProd };
+
+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());
+  }
+};
+
+template <ReduceType red_type, typename T>
+class CAllReduceOpCPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_THROW("CAllReduce op do not support CPUKernel for now.");
+  }
+};
+
+template <ReduceType red_type, typename T>
+class CAllReduceOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto place = ctx.GetPlace();
-    PADDLE_ENFORCE(is_gpu_place(place),
-                   "CAllReduce op can run on gpu place only for now.");
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     auto in = ctx.Input<framework::Tensor>("X");
     auto out = ctx.Output<framework::Tensor>("Out");
 
+    auto place = ctx.GetPlace();
     ncclDataType_t dtype = platform::ToNCCLDataType(in->type());
     int64_t numel = in->numel();
     const void* sendbuff = in->data<void>();
@@ -49,23 +72,6 @@ class CAllReduceOpKernel : public framework::OpKernel<T> {
     int rid = ctx.Attr<int>("ring_id");
     auto comm = platform::NCCLCommContext::Instance().Get(rid);
 
-    int reduce_type = ctx.Attr<int>("reduce_type");
-    ncclRedOp_t red_type = ncclSum;
-    switch (reduce_type) {
-      case 0:
-        red_type = ncclSum;
-        break;
-      case 1:
-        red_type = ncclProd;
-        break;
-      case 2:
-        red_type = ncclMax;
-        break;
-      case 3:
-        red_type = ncclMin;
-        break;
-    }
-
     cudaStream_t stream = nullptr;
     if (ctx.Attr<bool>("use_calc_stream")) {
       auto dev_ctx = platform::DeviceContextPool::Instance().Get(place);
@@ -74,13 +80,60 @@ class CAllReduceOpKernel : public framework::OpKernel<T> {
       stream = comm->stream();
     }
 
+    ncclRedOp_t nccl_red_type = ncclSum;
+    switch (red_type) {
+      case kRedSum:
+        nccl_red_type = ncclSum;
+        break;
+
+      case kRedMax:
+        nccl_red_type = ncclMax;
+        break;
+
+      case kRedMin:
+        nccl_red_type = ncclMin;
+        break;
+
+      case kRedProd:
+        nccl_red_type = ncclProd;
+        break;
+
+      default:
+        PADDLE_THROW("Invalid reduce type: %d", red_type);
+    }
+
     PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
-        sendbuff, recvbuff, numel, dtype, red_type, comm->comm(), stream));
+        sendbuff, recvbuff, numel, dtype, nccl_red_type, comm->comm(), stream));
 #else
     PADDLE_THROW("PaddlePaddle should compile with GPU.");
 #endif
   }
 };
 
+class CAllReduceOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "(Tensor), tensor to be allreduced.");
+    AddOutput("Out", "(Tensor) the allreduced result.");
+    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(string::Sprintf(R"DOC(
+CAllReduce %s Operator
+
+Call collective AllReduce with reduce type %s. If input and output are
+the same variable, in-place allreduce will be used.
+Reference: https://docs.nvidia.com/deeplearning/sdk/nccl-developer-guide/docs/usage/operations.html#allreduce
+)DOC",
+                               GetName(), GetName()));
+  }
+
+ protected:
+  virtual std::string GetName() const = 0;
+};
+
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/collective/c_allreduce_prod_op.cc

+/* 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>)

paddle/fluid/operators/collective/c_allreduce_prod_op.cu.cc

+/* 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>)

paddle/fluid/operators/collective/c_allreduce_op.cc → paddle/fluid/operators/collective/c_allreduce_sum_op.cc

@@ -12,58 +12,29 @@ 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 {
+class CAllReduceSumOpGradMaker : public framework::SingleGradOpDescMaker {
  public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
-  void InferShape(framework::InferShapeContext* ctx) const override {
-    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
-  }
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
 
  protected:
-  framework::OpKernelType GetExpectedKernelType(
-      const framework::ExecutionContext& ctx) const override {
-    return framework::OpKernelType(ctx.Input<framework::Tensor>("X")->type(),
-                                   ctx.GetPlace());
+  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 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");
-  }
+class CAllReduceSumOpMaker : public CAllReduceOpMaker {
+ protected:
+  std::string GetName() const override { return "Sum"; }
 };
 
 }  // namespace operators
@@ -72,12 +43,12 @@ If input and output are the same variable, in-place allreduce will be used.
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
-REGISTER_OP_WITHOUT_GRADIENT(c_allreduce, ops::CAllReduceOp,
-                             ops::CAllReduceOpMaker);
+REGISTER_OPERATOR(c_allreduce_sum, ops::CAllReduceOp,
+                  ops::CAllReduceSumOpGradMaker, ops::CAllReduceSumOpMaker);
 
-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>);
+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>)

paddle/fluid/operators/collective/c_allreduce_sum_op.cu.cc

+/* 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>)

paddle/fluid/operators/collective/c_broadcast_op.cc

@@ -12,9 +12,6 @@ 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_broadcast_op.h"
 
 namespace paddle {
@@ -50,9 +47,9 @@ class CBroadcastOpMaker : public framework::OpProtoAndCheckerMaker {
         "(bool default false) eject CUDA operations to calculation stream.")
         .SetDefault(false);
     AddComment(R"DOC(
-***CBroadcast Operator***
+CBroadcast Operator
 
-Call ncclBcast internally.
+Reference: https://docs.nvidia.com/deeplearning/sdk/nccl-developer-guide/docs/usage/operations.html#broadcast
 )DOC");
   }
 };
@@ -66,9 +63,8 @@ namespace plat = paddle::platform;
 REGISTER_OP_WITHOUT_GRADIENT(c_broadcast, ops::CBroadcastOp,
                              ops::CBroadcastOpMaker);
 
-REGISTER_OP_CPU_KERNEL(
-    c_broadcast, ops::CBroadcastOpKernel<plat::CPUDeviceContext, float>,
-    ops::CBroadcastOpKernel<plat::CPUDeviceContext, double>,
-    ops::CBroadcastOpKernel<plat::CPUDeviceContext, int>,
-    ops::CBroadcastOpKernel<plat::CPUDeviceContext, int64_t>,
-    ops::CBroadcastOpKernel<plat::CPUDeviceContext, plat::float16>);
+REGISTER_OP_CPU_KERNEL(c_broadcast, ops::CBroadcastOpCPUKernel<float>,
+                       ops::CBroadcastOpCPUKernel<double>,
+                       ops::CBroadcastOpCPUKernel<int>,
+                       ops::CBroadcastOpCPUKernel<int64_t>,
+                       ops::CBroadcastOpCPUKernel<plat::float16>);

paddle/fluid/operators/collective/c_broadcast_op.cu.cc

@@ -14,12 +14,74 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/collective/c_broadcast_op.h"
 
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#include "paddle/fluid/platform/collective_helper.h"
+#include "paddle/fluid/platform/nccl_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class CBroadcastOpCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+    auto x = ctx.Input<framework::LoDTensor>("X");
+    auto out = ctx.Output<framework::LoDTensor>("Out");
+    int numel = x->numel();
+    ncclDataType_t dtype = platform::ToNCCLDataType(x->type());
+
+    int rid = ctx.Attr<int>("ring_id");
+    auto comm = platform::NCCLCommContext::Instance().Get(rid);
+
+    auto place = ctx.GetPlace();
+    cudaStream_t stream = nullptr;
+    if (ctx.Attr<bool>("use_calc_stream")) {
+      auto dev_ctx = platform::DeviceContextPool::Instance().Get(place);
+      stream = static_cast<platform::CUDADeviceContext*>(dev_ctx)->stream();
+    } else {
+      stream = comm->stream();
+    }
+
+    int root = ctx.Attr<int>("root");
+    if (root == comm->rank()) {
+      PADDLE_ENFORCE(platform::dynload::ncclBcast(
+          reinterpret_cast<void*>(const_cast<T*>(x->data<T>())), numel, dtype,
+          root, comm->comm(), stream));
+      VLOG(3) << "rank " << comm->rank() << " invoke Bcast. sent "
+              << x->numel();
+
+      if (out != x) {
+        framework::TensorCopy(
+            *static_cast<const framework::Tensor*>(x), place,
+            *platform::DeviceContextPool::Instance().Get(place),
+            static_cast<framework::Tensor*>(out));
+      }
+    } else {
+      PADDLE_ENFORCE(platform::dynload::ncclBcast(out->mutable_data<T>(place),
+                                                  numel, dtype, root,
+                                                  comm->comm(), stream));
+      VLOG(3) << "rank " << comm->rank() << " invoke Bcast. recieved "
+              << framework::product(out->dims());
+    }
+
+    out->Resize(x->dims());
+    out->set_lod(x->lod());
+#else
+    PADDLE_THROW("PaddlePaddle should compile with GPU.");
+#endif
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
-REGISTER_OP_CUDA_KERNEL(
-    c_broadcast, ops::CBroadcastOpKernel<plat::CUDADeviceContext, float>,
-    ops::CBroadcastOpKernel<plat::CUDADeviceContext, double>,
-    ops::CBroadcastOpKernel<plat::CUDADeviceContext, int>,
-    ops::CBroadcastOpKernel<plat::CUDADeviceContext, int64_t>,
-    ops::CBroadcastOpKernel<plat::CUDADeviceContext, plat::float16>);
+REGISTER_OP_CUDA_KERNEL(c_broadcast, ops::CBroadcastOpCUDAKernel<float>,
+                        ops::CBroadcastOpCUDAKernel<double>,
+                        ops::CBroadcastOpCUDAKernel<int>,
+                        ops::CBroadcastOpCUDAKernel<int64_t>,
+                        ops::CBroadcastOpCUDAKernel<plat::float16>);

paddle/fluid/operators/collective/c_broadcast_op.h

@@ -22,69 +22,14 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-#include "paddle/fluid/platform/collective_helper.h"
-#include "paddle/fluid/platform/nccl_helper.h"
-#endif
-
 namespace paddle {
 namespace operators {
 
-template <typename DeviceContext, typename T>
-class CBroadcastOpKernel : public framework::OpKernel<T> {
+template <typename T>
+class CBroadcastOpCPUKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto place = ctx.GetPlace();
-    PADDLE_ENFORCE(is_gpu_place(place),
-                   "CBroadcastOp can run on gpu place only for now.");
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-    auto x = ctx.Input<framework::LoDTensor>("X");
-    auto out = ctx.Output<framework::LoDTensor>("Out");
-    int numel = x->numel();
-    ncclDataType_t dtype = platform::ToNCCLDataType(x->type());
-
-    int rid = ctx.Attr<int>("ring_id");
-    auto comm = platform::NCCLCommContext::Instance().Get(rid);
-
-    cudaStream_t stream = nullptr;
-    if (ctx.Attr<bool>("use_calc_stream")) {
-      auto dev_ctx = platform::DeviceContextPool::Instance().Get(place);
-      stream = static_cast<platform::CUDADeviceContext*>(dev_ctx)->stream();
-    } else {
-      stream = comm->stream();
-    }
-
-    int root = ctx.Attr<int>("root");
-    int nranks = comm->nranks();
-    PADDLE_ENFORCE(root >= 0 && root < nranks,
-                   "Expected root in range of [0,%d),but get %d", nranks, root);
-    if (root == comm->rank()) {
-      PADDLE_ENFORCE(platform::dynload::ncclBcast(
-          reinterpret_cast<void*>(const_cast<T*>(x->data<T>())), numel, dtype,
-          root, comm->comm(), stream));
-      VLOG(3) << "rank " << comm->rank() << " invoke Bcast. sent "
-              << x->numel();
-
-      if (out != x) {
-        // TODO(liuyi05): check inplace
-        framework::TensorCopy(
-            *static_cast<const framework::Tensor*>(x), place,
-            *platform::DeviceContextPool::Instance().Get(place),
-            static_cast<framework::Tensor*>(out));
-      }
-    } else {
-      PADDLE_ENFORCE(platform::dynload::ncclBcast(out->mutable_data<T>(place),
-                                                  numel, dtype, root,
-                                                  comm->comm(), stream));
-      VLOG(3) << "rank " << comm->rank() << " invoke Bcast. recieved "
-              << framework::product(out->dims());
-    }
-
-    out->Resize(x->dims());
-    out->set_lod(x->lod());
-#else
-    PADDLE_THROW("PaddlePaddle should compile with GPU.");
-#endif
+    PADDLE_THROW("Unimplemented cpu kernel for CBroadcastOp.");
   }
 };
 

paddle/fluid/operators/collective/c_comm_init_op.cc

@@ -28,6 +28,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/collective_helper.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
+
 namespace paddle {
 namespace operators {
 

paddle/fluid/operators/collective/c_gen_nccl_id_op.cc

@@ -11,9 +11,11 @@ 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. */
+
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include <nccl.h>
 #endif
+
 #include <stdint.h>
 #include <ostream>
 #include <string>
@@ -24,9 +26,11 @@ limitations under the License. */
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/operators/distributed/distributed.h"
 #include "paddle/fluid/operators/distributed/request_handler_impl.h"
+
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
+
 namespace paddle {
 namespace operators {
 

paddle/fluid/operators/collective/c_reducescatter_op.cc

@@ -13,9 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/collective/c_reducescatter_op.h"
-#include <future>  // NOLINT
+
 #include <memory>
-#include <ostream>
 
 namespace paddle {
 namespace operators {
@@ -54,9 +53,9 @@ class CReduceScatterOpMaker : public framework::OpProtoAndCheckerMaker {
         "(bool default false) eject CUDA operations to calculation stream.")
         .SetDefault(false);
     AddComment(R"DOC(
-***CReduceScatter Operator***
+CReduceScatter Operator
 
-Call NCCL collective ReduceScatter internally.
+Reference: https://docs.nvidia.com/deeplearning/sdk/nccl-developer-guide/docs/usage/operations.html#reducescatter
 )DOC");
   }
 };
@@ -85,9 +84,8 @@ namespace plat = paddle::platform;
 REGISTER_OPERATOR(c_reducescatter, ops::CReduceScatterOp,
                   ops::CReduceScatterOpMaker);
 
-REGISTER_OP_CPU_KERNEL(
-    c_reducescatter, ops::CReduceScatterOpKernel<plat::CPUDeviceContext, float>,
-    ops::CReduceScatterOpKernel<plat::CPUDeviceContext, double>,
-    ops::CReduceScatterOpKernel<plat::CPUDeviceContext, int>,
-    ops::CReduceScatterOpKernel<plat::CPUDeviceContext, int64_t>,
-    ops::CReduceScatterOpKernel<plat::CPUDeviceContext, plat::float16>);
+REGISTER_OP_CPU_KERNEL(c_reducescatter, ops::CReduceScatterOpCPUKernel<float>,
+                       ops::CReduceScatterOpCPUKernel<double>,
+                       ops::CReduceScatterOpCPUKernel<int>,
+                       ops::CReduceScatterOpCPUKernel<int64_t>,
+                       ops::CReduceScatterOpCPUKernel<plat::float16>);

paddle/fluid/operators/collective/c_reducescatter_op.cu.cc

@@ -14,13 +14,61 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/collective/c_reducescatter_op.h"
 
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#include "paddle/fluid/platform/collective_helper.h"
+#include "paddle/fluid/platform/nccl_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class CReduceScatterOpCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+    auto in = ctx.Input<framework::Tensor>("X");
+    auto out = ctx.Output<framework::Tensor>("Out");
+
+    int rid = ctx.Attr<int>("ring_id");
+    auto comm = platform::NCCLCommContext::Instance().Get(rid);
+    int nranks = comm->nranks();
+
+    auto place = ctx.GetPlace();
+    auto out_dims = in->dims();
+    out_dims[0] = out_dims[0] / nranks;
+    out->mutable_data<T>(out_dims, place);
+
+    int64_t recv_numel = in->numel() / nranks;
+    const T* send_buff = in->data<T>();
+    T* recv_buff = out->data<T>();
+    int dtype = platform::ToNCCLDataType(in->type());
+
+    cudaStream_t stream = nullptr;
+    if (ctx.Attr<bool>("use_calc_stream")) {
+      auto dev_ctx = platform::DeviceContextPool::Instance().Get(place);
+      stream = static_cast<platform::CUDADeviceContext*>(dev_ctx)->stream();
+    } else {
+      stream = comm->stream();
+    }
+
+    PADDLE_ENFORCE(platform::dynload::ncclReduceScatter(
+        send_buff, recv_buff, recv_numel, static_cast<ncclDataType_t>(dtype),
+        ncclSum, comm->comm(), stream));
+#else
+    PADDLE_THROW("PaddlePaddle should compile with GPU.");
+#endif
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
-REGISTER_OP_CUDA_KERNEL(
-    c_reducescatter,
-    ops::CReduceScatterOpKernel<plat::CUDADeviceContext, float>,
-    ops::CReduceScatterOpKernel<plat::CUDADeviceContext, double>,
-    ops::CReduceScatterOpKernel<plat::CUDADeviceContext, int>,
-    ops::CReduceScatterOpKernel<plat::CUDADeviceContext, int64_t>,
-    ops::CReduceScatterOpKernel<plat::CUDADeviceContext, plat::float16>);
+REGISTER_OP_CUDA_KERNEL(c_reducescatter, ops::CReduceScatterOpCUDAKernel<float>,
+                        ops::CReduceScatterOpCUDAKernel<double>,
+                        ops::CReduceScatterOpCUDAKernel<int>,
+                        ops::CReduceScatterOpCUDAKernel<int64_t>,
+                        ops::CReduceScatterOpCUDAKernel<plat::float16>);

paddle/fluid/operators/collective/c_reducescatter_op.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+
 #include <algorithm>
 #include <utility>
 #include <vector>
@@ -22,52 +23,14 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-#include "paddle/fluid/platform/collective_helper.h"
-#include "paddle/fluid/platform/nccl_helper.h"
-#endif
-
 namespace paddle {
 namespace operators {
 
-template <typename DeviceContext, typename T>
-class CReduceScatterOpKernel : public framework::OpKernel<T> {
+template <typename T>
+class CReduceScatterOpCPUKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto place = ctx.GetPlace();
-    PADDLE_ENFORCE(is_gpu_place(place),
-                   "CAllReduce op can run on gpu place only for now.");
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-    auto in = ctx.Input<framework::Tensor>("X");
-    auto out = ctx.Output<framework::Tensor>("Out");
-
-    int rid = ctx.Attr<int>("ring_id");
-    auto comm = platform::NCCLCommContext::Instance().Get(rid);
-    int nranks = comm->nranks();
-
-    auto out_dims = in->dims();
-    out_dims[0] = out_dims[0] / nranks;
-    out->mutable_data<T>(out_dims, place);
-
-    int64_t recv_numel = in->numel() / nranks;
-    const T* send_buff = in->data<T>();
-    T* recv_buff = out->data<T>();
-    int dtype = platform::ToNCCLDataType(in->type());
-
-    cudaStream_t stream = nullptr;
-    if (ctx.Attr<bool>("use_calc_stream")) {
-      auto dev_ctx = platform::DeviceContextPool::Instance().Get(place);
-      stream = static_cast<platform::CUDADeviceContext*>(dev_ctx)->stream();
-    } else {
-      stream = comm->stream();
-    }
-
-    PADDLE_ENFORCE(platform::dynload::ncclReduceScatter(
-        send_buff, recv_buff, recv_numel, static_cast<ncclDataType_t>(dtype),
-        ncclSum, comm->comm(), stream));
-#else
-    PADDLE_THROW("PaddlePaddle should compile with GPU.");
-#endif
+    PADDLE_THROW("Unimplemented cpu kernel for CReduceScatterOp.");
   }
 };
 

paddle/fluid/operators/collective/c_sync_calc_stream_op.cc

@@ -15,12 +15,12 @@ limitations under the License. */
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include <nccl.h>
 #endif
-#include <stdint.h>
-#include <ostream>
+
 #include <string>
 
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
+
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/collective_helper.h"
 #endif
@@ -40,7 +40,6 @@ class CSyncCalcStreamOp : public framework::OperatorBase {
                const platform::Place& place) const override {
     PADDLE_ENFORCE(is_gpu_place(place),
                    "Sync stream op can run on gpu place only for now.");
-
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     auto dev_ctx = static_cast<platform::CUDADeviceContext*>(
         platform::DeviceContextPool::Instance().Get(place));
@@ -57,12 +56,12 @@ class CSyncCalcStreamOp : public framework::OperatorBase {
 class CSyncCalcStreamOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {
-    AddInput("X", "(Tensor) Dependency of last param need to sync");
-    AddOutput("Out", "(Tensor) Dependency of last param need to sync");
+    AddInput("X", "(Tensor) Dependency of the variable need to sync");
+    AddOutput("Out", "(Tensor) Dependency of the variable need to sync");
     AddComment(R"DOC(
-***Sync Operator***
+CSyncCalcStream Operator
 
-Call cuda stream synchronize.
+Call calculation stream synchronization.
 )DOC");
   }
 };

paddle/fluid/operators/collective/c_sync_comm_stream_op.cc

@@ -11,11 +11,11 @@ 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. */
+
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include <nccl.h>
 #endif
-#include <stdint.h>
-#include <ostream>
+
 #include <string>
 
 #include "paddle/fluid/framework/lod_tensor.h"
@@ -57,13 +57,13 @@ class CSyncCommStreamOp : public framework::OperatorBase {
 class CSyncCommStreamOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {
-    AddInput("X", "(Tensor) Dependency of last param need to sync");
-    AddOutput("Out", "(Tensor) Dependency of last param need to sync");
+    AddInput("X", "(Tensor) Dependency of the variable need to sync");
+    AddOutput("Out", "(Tensor) Dependency of the variable need to sync");
     AddAttr<int>("ring_id", "(int default 0) ring id.").SetDefault(0);
     AddComment(R"DOC(
-***Sync Operator***
+CSyncCommStream Operator
 
-Call nccl stream synchronize.
+Call communication stream synchronization.
 )DOC");
   }
 };

paddle/fluid/platform/collective_helper.cc

@@ -12,11 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-// #ifndef _WIN32
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/collective_helper.h"
 
-#include <functional>
+#include <memory>
+#include <utility>
 
 #include "paddle/fluid/platform/dynload/nccl.h"
 
@@ -34,24 +34,23 @@ class NCCLCommImpl : public NCCLComm {
   void set_rank(int rank) { rank_ = rank; }
   int rank() const override { return rank_; }
 
-  void set_local_rank(int local_rank) { local_rank_ = local_rank; }
-  int local_rank() const override { return local_rank_; }
-
-  void set_comm(ncclComm_t comm) { comm_ = comm; }
-  ncclComm_t comm() const override { return comm_; }
+  int device_id() const override {
+    return boost::get<CUDAPlace>(dev_ctx_->GetPlace()).device;
+  }
 
-  void set_dev_ctx(CUDADeviceContext* dev_ctx) { dev_ctx_ = dev_ctx; }
-  CUDADeviceContext* DevCtx() const override { return dev_ctx_; }
+  ncclComm_t comm() const override { return dev_ctx_->nccl_comm(); }
 
   cudaStream_t stream() const override { return dev_ctx_->stream(); }
 
+  void set_dev_ctx(std::unique_ptr<CUDADeviceContext>&& dev_ctx) {
+    dev_ctx_ = std::move(dev_ctx);
+  }
+
  private:
   int ring_id_;
   int nranks_;
   int rank_;
-  int local_rank_;
-  ncclComm_t comm_;
-  CUDADeviceContext* dev_ctx_;
+  std::unique_ptr<CUDADeviceContext> dev_ctx_;
 };
 
 // NOTE: not thread-safe
@@ -73,13 +72,15 @@ NCCLComm* NCCLCommContext::CreateNCCLComm(ncclUniqueId* nccl_id, int nranks,
   PADDLE_ENFORCE(
       platform::dynload::ncclCommInitRank(&comm, nranks, *nccl_id, rank));
 
+  std::unique_ptr<CUDADeviceContext> dev_ctx(
+      new CUDADeviceContext(CUDAPlace(dev_id)));
+  dev_ctx->set_nccl_comm(comm);
+
   NCCLCommImpl* communicator = new NCCLCommImpl;
   communicator->set_ring_id(ring_id);
   communicator->set_nranks(nranks);
   communicator->set_rank(rank);
-  communicator->set_local_rank(dev_id);
-  communicator->set_comm(comm);
-  communicator->set_dev_ctx(dev_ctx_map_.at(dev_id).get());
+  communicator->set_dev_ctx(std::move(dev_ctx));
 
   comm_map_.emplace(ring_id, std::unique_ptr<NCCLComm>(communicator));
 

paddle/fluid/platform/collective_helper.h

@@ -12,10 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-// #ifndef _WIN32
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #pragma once
 
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include <memory>
 #include <string>
 #include <unordered_map>
@@ -53,10 +52,9 @@ class NCCLComm {
   virtual int ring_id() const = 0;
   virtual int nranks() const = 0;
   virtual int rank() const = 0;
-  virtual int local_rank() const = 0;
+  virtual int device_id() const = 0;
   virtual ncclComm_t comm() const = 0;
   virtual cudaStream_t stream() const = 0;
-  virtual CUDADeviceContext* DevCtx() const = 0;
   virtual ~NCCLComm() = default;
 };
 
@@ -73,16 +71,6 @@ class NCCLCommContext {
   NCCLComm* CreateNCCLComm(ncclUniqueId* nccl_id, int nranks, int rank,
                            int dev_id, int ring_id = 0);
 
-  CUDADeviceContext* DevCtx(int dev_id) const {
-    PADDLE_ENFORCE(dev_ctx_map_.count(dev_id),
-                   "CUDADeviceContext at device %d has not been initialized");
-    return dev_ctx_map_.at(dev_id).get();
-  }
-
-  CUDADeviceContext* DevCtx(platform::Place p) const {
-    return DevCtx(boost::get<CUDAPlace>(p).device);
-  }
-
   // retrieve a communicator by the ring id
   NCCLComm* Get(int ring_id) const {
     PADDLE_ENFORCE(comm_map_.count(ring_id),

paddle/fluid/pybind/const_value.cc

@@ -46,7 +46,6 @@ void BindConstValue(pybind11::module* m) {
       .value("Loss", framework::OpRole::kLoss)
       .value("RPC", framework::OpRole::kRPC)
       .value("Dist", framework::OpRole::kDist)
-      .value("Collective", framework::OpRole::kCollective)
       .value("LRSched", framework::OpRole::kLRSched);
 
   op_proto_and_checker_maker.def(

python/paddle/fluid/tests/unittests/collective_allreduce_op.py

@@ -42,7 +42,6 @@ class TestCollectiveAllreduce(TestCollectiveRunnerBase):
 
     def get_model(self, main_prog, startup_program):
         ring_id = 0
-        reduce_type = 0
         with fluid.program_guard(main_prog, startup_program):
             tindata = layers.data(
                 name="tindata", shape=[10, 1000], dtype='float32')
@@ -53,10 +52,9 @@ class TestCollectiveAllreduce(TestCollectiveRunnerBase):
                 persistable=False,
                 stop_gradient=False)
             main_prog.global_block().append_op(
-                type="c_allreduce",
+                type="c_allreduce_sum",
                 inputs={'X': tindata},
-                attrs={'ring_id': ring_id,
-                       'reduce_type': reduce_type},
+                attrs={'ring_id': ring_id},
                 outputs={'Out': toutdata})
             main_prog.global_block().append_op(
                 type="c_sync_comm_stream",

python/paddle/fluid/transpiler/collective.py

-# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#   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.
@@ -37,8 +37,8 @@ class Collective(object):
     '''
     '''
 
-    def __init__(self):
-        self.global_ring_id = 0
+    def __init__(self, nrings):
+        self.nrings = nrings
         self.endpoints = None
         self.current_endpoint = None
         self.nranks = None
@@ -90,9 +90,10 @@ class Collective(object):
         raise NotImplementedError('call the inherited method of subclasses')
 
     def _transpile_startup_program(self):
-        self._init_communicator(self.startup_program, self.current_endpoint,
-                                self.endpoints, self.rank, self.global_ring_id,
-                                self.wait_port)
+        for ring_id in range(self.nrings):
+            self._init_communicator(self.startup_program, self.current_endpoint,
+                                    self.endpoints, self.rank, ring_id,
+                                    self.wait_port)
         self._broadcast_params()
 
     def _init_communicator(self, program, current_endpoint, endpoints, rank,
@@ -116,7 +117,7 @@ class Collective(object):
                 'rank': rank,
                 'endpoint': current_endpoint,
                 'other_endpoints': other_endpoints,
-                self.op_role_key: OpRole.Collective
+                self.op_role_key: OpRole.Forward
             })
         block.append_op(
             type='c_comm_init',
@@ -126,29 +127,31 @@ class Collective(object):
                 'nranks': nranks,
                 'rank': rank,
                 'ring_id': ring_id,
-                self.op_role_key: OpRole.Collective
+                self.op_role_key: OpRole.Forward
             })
 
     def _broadcast_params(self):
         block = self.startup_program.global_block()
-        for var in block.iter_parameters():
+        ring_id = -1
+        for param in block.iter_parameters():
+            ring_id = (ring_id + 1) % self.nrings
             block.append_op(
                 type='c_broadcast',
-                inputs={'X': var},
-                outputs={'Out': var},
+                inputs={'X': param},
+                outputs={'Out': param},
                 attrs={
-                    'ring_id': self.global_ring_id,
+                    'ring_id': ring_id,
                     'root': 0,
-                    self.op_role_key: OpRole.Collective
+                    self.op_role_key: OpRole.Forward
                 })
-        block.append_op(
-            type='c_sync_comm_stream',
-            inputs={'X': var},
-            outputs={'Out': var},
-            attrs={
-                'ring_id': self.global_ring_id,
-                self.op_role_key: OpRole.Collective
-            })
+
+        for ring_id in range(self.nrings):
+            block.append_op(
+                type='c_sync_comm_stream',
+                inputs={'X': param},
+                outputs={'Out': param},
+                attrs={'ring_id': ring_id,
+                       self.op_role_key: OpRole.Forward})
 
     def _is_loss_grad_op(self, op):
         if self.op_role_key not in op.attr_names:
@@ -173,8 +176,8 @@ class GradAllReduce(Collective):
     '''
     '''
 
-    def __init__(self):
-        Collective.__init__(self)
+    def __init__(self, nrings=2):
+        Collective.__init__(self, nrings)
 
     def _transpile_main_program(self):
         self._insert_scale_loss_grad_ops()
@@ -196,11 +199,13 @@ class GradAllReduce(Collective):
                     outputs={'Out': loss_grad_var},
                     attrs={
                         'scale': 1.0 / self.nranks,
-                        self.op_role_key: OpRole.Collective
+                        self.op_role_key: OpRole.Backward
                     })
 
     def _insert_allreduce_ops(self):
         block = self.main_program.global_block()
+        ring_id = -1
+        grad = None
         for idx, op in reversed(list(enumerate(block.ops))):
             if self._is_backward_op(op) and \
                     self.op_role_var_key in op.attr_names:
@@ -208,41 +213,50 @@ class GradAllReduce(Collective):
 
                 if len(op_role_var) == 0:
                     continue
-
                 assert len(op_role_var) % 2 == 0
 
-                block._insert_op(
-                    idx + 1,
-                    type='c_sync_calc_stream',
-                    inputs={'X': block.vars[grad]},
-                    outputs={'Out': block.vars[grad]},
-                    attrs={self.op_role_key: OpRole.Collective})
-
-                offset = 2
+                offset = idx
                 for i in range(0, len(op_role_var), 2):
-                    grad = op_role_var[i + 1]
+                    param = block.vars[op_role_var[i]]
+                    grad = block.vars[op_role_var[i + 1]]
+                    if offset == idx:
+                        offset += 1
+                        block._insert_op(
+                            offset,
+                            type='c_sync_calc_stream',
+                            inputs={'X': grad},
+                            outputs={'Out': grad},
+                            attrs={self.op_role_key: OpRole.Backward})
+                        offset += 1
+
+                    # As we search ops reversedly, we should insert c_allreduce_sum
+                    # op in the same way to keep the ring_id alternate
+                    ring_id = (ring_id + 1) % self.nrings
                     block._insert_op(
-                        idx + offset,
-                        type='c_allreduce',
-                        inputs={'X': [block.vars[grad]]},
-                        outputs={'Out': [block.vars[grad]]},
+                        offset,
+                        type='c_allreduce_sum',
+                        inputs={'X': grad},
+                        outputs={'Out': grad},
                         attrs={
-                            'reduce_type': 0,
-                            self.op_role_key: OpRole.Collective
+                            'ring_id': ring_id,
+                            self.op_role_key: OpRole.Backward
                         })
-                    offset += 1
+
+        if grad is None:
+            return
 
         for idx, op in enumerate(block.ops):
             if self._is_optimizer_op(op):
-                block._insert_op(
-                    idx,
-                    type='c_sync_comm_stream',
-                    inputs={'X': block.vars[grad]},
-                    outputs={'Out': block.vars[grad]},
-                    attrs={
-                        'ring_id': self.global_ring_id,
-                        self.op_role_key: OpRole.Collective
-                    })
+                for ring_id in range(self.nrings):
+                    block._insert_op(
+                        idx + ring_id,
+                        type='c_sync_comm_stream',
+                        inputs={'X': grad},
+                        outputs={'Out': grad},
+                        attrs={
+                            'ring_id': ring_id,
+                            self.op_role_key: OpRole.Backward
+                        })
                 break
 
 
@@ -250,8 +264,8 @@ class LocalSGD(Collective):
     '''
     '''
 
-    def __init__(self):
-        Collective.__init__(self)
+    def __init__(self, nrings=2):
+        Collective.__init__(self, nrings)
         self.snapshot_key = '@SNAPSHOT'
 
     def _transpile_startup_program(self):
@@ -268,7 +282,7 @@ class LocalSGD(Collective):
                 type='assign',
                 inputs={'X': [param]},
                 outputs={'Out': [snapshot]},
-                attrs={self.op_role_key: OpRole.Collective})
+                attrs={self.op_role_key: OpRole.Forward})
 
     def snapshot_name(self, param_name):
         return param_name + self.snapshot_key
@@ -276,6 +290,7 @@ class LocalSGD(Collective):
     def _transpile_main_program(self):
         block = self.main_program.global_block()
         ordered_param_snapshot = []
+        ring_id = -1
         for idx, op in reversed(list(enumerate(block.ops))):
             if self._is_update_op(op):
                 param = block.vars[op.input('Param')[0]]
@@ -291,33 +306,33 @@ class LocalSGD(Collective):
                     inputs={'X': [snapshot],
                             'Y': [param]},
                     outputs={'Out': [param]},
-                    attrs={self.op_role_key: OpRole.Collective})
+                    attrs={self.op_role_key: OpRole.Optimize})
                 block._insert_op(
                     idx + 2,
                     type='c_sync_calc_stream',
                     inputs={'X': param},
                     outputs={'Out': param},
-                    attrs={self.op_role_key: OpRole.Collective})
+                    attrs={self.op_role_key: OpRole.Optimize})
+                ring_id = (ring_id + 1) % self.nrings
                 block._insert_op(
                     idx + 3,
-                    type='c_allreduce',
+                    type='c_allreduce_sum',
                     inputs={'X': [param]},
                     outputs={'Out': [param]},
                     attrs={
-                        'reduce_type': 0,
-                        self.op_role_key: OpRole.Collective
+                        'ring_id': ring_id,
+                        self.op_role_key: OpRole.Optimize
                     })
 
                 ordered_param_snapshot.append((param, snapshot))
 
-        block.append_op(
-            type='c_sync_comm_stream',
-            inputs={'X': param},
-            outputs={'Out': param},
-            attrs={
-                'ring_id': self.global_ring_id,
-                self.op_role_key: OpRole.Collective
-            })
+        for ring_id in range(self.nrings):
+            block.append_op(
+                type='c_sync_comm_stream',
+                inputs={'X': param},
+                outputs={'Out': param},
+                attrs={'ring_id': ring_id,
+                       self.op_role_key: OpRole.Optimize})
 
         for param_snapshot in reversed(ordered_param_snapshot):
             param = param_snapshot[0]
@@ -328,16 +343,16 @@ class LocalSGD(Collective):
                 outputs={'Out': [param]},
                 attrs={
                     'scale': 1.0 / self.nranks,
-                    self.op_role_key: OpRole.Collective
+                    self.op_role_key: OpRole.Optimize
                 })
             block.append_op(
                 type='elementwise_sub',
                 inputs={'X': [snapshot],
                         'Y': [param]},
                 outputs={'Out': [param]},
-                attrs={self.op_role_key: OpRole.Collective})
+                attrs={self.op_role_key: OpRole.Optimize})
             block.append_op(
                 type='assign',
                 inputs={'X': [param]},
                 outputs={'Out': [snapshot]},
-                attrs={self.op_role_key: OpRole.Collective})
+                attrs={self.op_role_key: OpRole.Optimize})