add gather op

paddle/fluid/operators/collective/c_gather_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_gather_op.h"
+
+namespace paddle {
+namespace operators {
+
+class CGatherOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "CGather");
+    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "CGather");
+    int root_id = ctx->Attrs().Get<int>("root");
+    int ring_id = ctx->Attrs().Get<int>("ring_id");
+    int nranks = ctx->Attrs().Get<int>("nranks");
+    PADDLE_ENFORCE_GE(nranks, 2,
+                      platform::errors::InvalidArgument(
+                          "The number of ranks (%d) must be greater than 1 "
+                          "to use collective op (c_gather op).",
+                          nranks));
+    PADDLE_ENFORCE_GE(
+        root_id, 0,
+        platform::errors::InvalidArgument(
+            "The root_id (%d) for c_gather_op must be non-negative.", root_id));
+    PADDLE_ENFORCE_LT(
+        root_id, nranks,
+        platform::errors::InvalidArgument(
+            "The root_id (%d) for c_gather_op must be less than nranks (%d).",
+            root_id, nranks));
+    PADDLE_ENFORCE_GE(
+        ring_id, 0,
+        platform::errors::InvalidArgument(
+            "The ring_id (%d) for c_gather_op must be non-negative.", root_id));
+    framework::DDim dim = ctx->GetInputDim("X");
+    dim[0] = dim[0] * nranks;
+    if (dim[0] < 0) dim[0] = -1;
+    ctx->SetOutputDim("Out", dim);
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        OperatorWithKernel::IndicateVarDataType(ctx, "X"), ctx.GetPlace());
+  }
+};
+
+class CGatherOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "(Tensor) tensor to be gathered.");
+    AddOutput("Out", "(Tensor) the result of gather.");
+    AddAttr<int>("ring_id", "(int default 0) nccl communication ring id.")
+        .SetDefault(0);
+    AddAttr<int>("root", "(int default 0) root id for broadcasting.")
+        .SetDefault(0);
+    AddAttr<int>("nranks", "(int default 1) number of ranks.").SetDefault(0);
+    AddAttr<bool>(
+        "use_calc_stream",
+        "(bool default false) eject CUDA operations to calculation stream.")
+        .SetDefault(false);
+    AddComment(R"DOC(
+CGather Operator
+Gather tensors from all participators.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
+REGISTER_OP_WITHOUT_GRADIENT(c_gather, ops::CGatherOp, ops::CGatherOpMaker);
+
+REGISTER_OP_CPU_KERNEL(c_gather, ops::CGatherOpCPUKernel<float>,
+                       ops::CGatherOpCPUKernel<double>,
+                       ops::CGatherOpCPUKernel<int>,
+                       ops::CGatherOpCPUKernel<int64_t>,
+                       ops::CGatherOpCPUKernel<plat::float16>);

paddle/fluid/operators/collective/c_gather_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_gather_op.h"
+
+#if defined(PADDLE_WITH_NCCL)
+#include "paddle/fluid/platform/collective_helper.h"
+#include "paddle/fluid/platform/nccl_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class CGatherOpCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+#if defined(PADDLE_WITH_NCCL)
+    auto x = ctx.Input<framework::LoDTensor>("X");
+    auto out = ctx.Output<framework::LoDTensor>("Out");
+    int send_numel = x->numel();
+    ncclDataType_t dtype = platform::ToNCCLDataType(x->type());
+
+    int nranks = ctx.Attr<int>("nranks");
+    int root_id = ctx.Attr<int>("root");
+    int ring_id = ctx.Attr<int>("ring_id");
+    auto place = ctx.GetPlace();
+    auto comm = platform::NCCLCommContext::Instance().Get(ring_id, place);
+    PADDLE_ENFORCE_EQ(nranks, comm->nranks(),
+                      platform::errors::InvalidArgument(
+                          "The number of ranks (%d) you set of must "
+                          "be equal to comm->nranks (%d).",
+                          nranks, comm->nranks()));
+    PADDLE_ENFORCE_GE(
+        root_id, 0,
+        platform::errors::InvalidArgument(
+            "The root_id (%d) for c_scatter_op must be non-negative.",
+            root_id));
+    PADDLE_ENFORCE_GE(
+        ring_id, 0,
+        platform::errors::InvalidArgument(
+            "The ring_id (%d) for c_scatter_op must be non-negative.",
+            ring_id));
+
+    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();
+    }
+
+    framework::DDim x_dims = x->dims();
+    framework::DDim out_dims(x_dims);
+    out_dims[0] *= nranks;
+    auto send_buf = x->data<T>();
+    auto offset = 0;
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclGroupStart());
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclSend(
+        send_buf, send_numel, dtype, root_id, comm->comm(), stream));
+    if (root_id == comm->rank()) {
+      auto recv_buf = out->mutable_data<T>(out_dims, place);
+      for (auto i = 0; i < nranks; ++i) {
+        PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclRecv(
+            recv_buf + offset, send_numel, dtype, i, comm->comm(), stream));
+        offset += send_numel;
+      }
+    }
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclGroupEnd());
+#else
+    PADDLE_ENFORCE_EQ(
+        true, false,
+        platform::errors::Unavailable("PaddlePaddle should compile with GPU."));
+#endif
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
+REGISTER_OP_CUDA_KERNEL(c_gather, ops::CGatherOpCUDAKernel<float>,
+                        ops::CGatherOpCUDAKernel<double>,
+                        ops::CGatherOpCUDAKernel<int>,
+                        ops::CGatherOpCUDAKernel<int64_t>,
+                        ops::CGatherOpCUDAKernel<plat::float16>);

paddle/fluid/operators/collective/c_gather_op.h

+/* 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 <algorithm>
+#include <utility>
+#include <vector>
+
+#include "paddle/fluid/framework/data_type.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+#if defined(PADDLE_WITH_GLOO)
+#include <gloo/gather.h>
+#include "paddle/fluid/framework/fleet/gloo_wrapper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class CGatherOpCPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+#if defined(PADDLE_WITH_GLOO)
+    auto in = ctx.Input<framework::Tensor>("X");
+    auto out = ctx.Output<framework::Tensor>("Out");
+    auto root_id = ctx.Attr<int>("root");
+    auto nranks = ctx.Attr<int>("nranks");
+
+    auto gloo = paddle::framework::GlooWrapper::GetInstance();
+    PADDLE_ENFORCE_EQ(
+        gloo->IsInitialized(), true,
+        platform::errors::PreconditionNotMet(
+            "You must initialize the gloo environment first to use it."));
+
+    PADDLE_ENFORCE_EQ(nranks, gloo->Size(),
+                      platform::errors::InvalidArgument(
+                          "The number of ranks (%d) you set must "
+                          "be equal to gloo->Size() (%d).",
+                          nranks, gloo->Size()));
+    int64_t send_numel = in->numel();
+    int64_t recv_numel = out->numel();
+    auto in_dim = x->dims();
+    auto out_dim = framework::DDim(in_dim);
+    out_dim[0] *= nranks;
+    auto nranks = gloo->Size();
+    auto rank = gloo->Rank();
+    gloo::GatherOptions opts(gloo->GetContext());
+    if (root_id == rank) {
+      T* recv_buff = out->mutable_data<T>(place, out_dim);
+      opts.setOutput(recv_buff, recv_numel);
+    }
+    opts.setInput(send_buff, send_numel);
+    opts.setRoot(root_id);
+
+    gloo::gather(opts);
+#else
+    PADDLE_THROW(platform::errors::Unavailable(
+        "PaddlePaddle should compile with GLOO by setting WITH_GLOO=ON"));
+#endif
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/collective/c_recv_op.cc

@@ -21,20 +21,34 @@ class CRecvOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
-  void InferShape(framework::InferShapeContext* ctx) const override {}
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "CRecv");
+    int peer = ctx->Attrs().Get<int>("peer");
+    int ring_id = ctx->Attrs().Get<int>("ring_id");
+    PADDLE_ENFORCE_GE(
+        peer, 0,
+        platform::errors::InvalidArgument(
+            "The peer (%d) for c_send_op must be non-negative.", peer));
+    PADDLE_ENFORCE_GE(
+        ring_id, 0,
+        platform::errors::InvalidArgument(
+            "The ring_id (%d) for c_send_op must be non-negative.", ring_id));
+  }
 
  protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
-    return framework::OpKernelType(
-        OperatorWithKernel::IndicateVarDataType(ctx, "Out"), ctx.GetPlace());
+    auto out = ctx.Output<framework::LoDTensor>("Out");
+    auto dtype = out->type();
+    return framework::OpKernelType(dtype, ctx.GetPlace());
+    // OperatorWithKernel::IndicateVarDataType(ctx, "Out"), ctx.GetPlace());
   }
 };
 
 class CRecvOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {
-    AddInput("Out", "(Tensor) tensor to receive.");
+    AddOutput("Out", "(Tensor) tensor to receive.");
     AddAttr<int>("ring_id", "(int default 0) nccl communication ring id.")
         .SetDefault(0);
     AddAttr<int>("peer", "(int default 0) rank id for sender.").SetDefault(0);

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

@@ -27,7 +27,7 @@ class CRecvOpCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
 #if defined(PADDLE_WITH_NCCL)
-    auto out = ctx.Input<framework::LoDTensor>("Out");
+    auto out = ctx.Output<framework::LoDTensor>("Out");
     int numel = out->numel();
     ncclDataType_t dtype = platform::ToNCCLDataType(out->type());
 
@@ -44,9 +44,13 @@ class CRecvOpCUDAKernel : public framework::OpKernel<T> {
     }
 
     int peer = ctx.Attr<int>("peer");
-    PADDLE_ENFORCE_CUDA_SUCCESS(
-        platform::dynload::ncclRecv(const_cast<T*>(out->data<T>()), numel,
-                                    dtype, peer, comm->comm(), stream));
+    PADDLE_ENFORCE_LT(
+        peer, comm->nranks(),
+        platform::errors::InvalidArgument("The value of peer (%d) you set must "
+                                          "be less than comm->nranks (%d).",
+                                          peer, comm->nranks()));
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclRecv(
+        out->data<T>(), numel, dtype, peer, comm->comm(), stream));
     VLOG(3) << "rank " << comm->rank() << " recv "
             << framework::product(out->dims()) << " from " << peer;
 #else

paddle/fluid/operators/collective/c_scatter_op.cc

@@ -37,11 +37,16 @@ class CScatterOp : public framework::OperatorWithKernel {
         platform::errors::InvalidArgument(
             "The root_id (%d) for c_scatter_op must be non-negative.",
             root_id));
+    PADDLE_ENFORCE_LT(root_id, nranks,
+                      platform::errors::InvalidArgument(
+                          "The root_id (%d) for c_scatter_op must be less "
+                          "than the number of ranks (%d).",
+                          root_id, nranks));
     PADDLE_ENFORCE_GE(
         ring_id, 0,
         platform::errors::InvalidArgument(
             "The ring_id (%d) for c_scatter_op must be non-negative.",
-            root_id));
+            ring_id));
     framework::DDim dim = ctx->GetInputDim("X");
     dim[0] = dim[0] / nranks;
     if (dim[0] < 0) dim[0] = -1;

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

@@ -39,7 +39,7 @@ class CScatterOpCUDAKernel : public framework::OpKernel<T> {
     auto comm = platform::NCCLCommContext::Instance().Get(ring_id, place);
     PADDLE_ENFORCE_EQ(nranks, comm->nranks(),
                       platform::errors::InvalidArgument(
-                          "The number of ranks (%d) you set of must "
+                          "The number of ranks (%d) you set must "
                           "be equal to comm->nranks (%d).",
                           nranks, comm->nranks()));
     PADDLE_ENFORCE_GE(
@@ -63,30 +63,23 @@ class CScatterOpCUDAKernel : public framework::OpKernel<T> {
 
     framework::DDim x_dims = x->dims();
     framework::DDim out_dims(x_dims);
-    framework::Tensor temp;
-    auto out_ptr = temp.mutable_data<T>(out_dims, place);
+    out_dims[0] /= nranks;
+    auto send_buf = x->data<T>();
+    auto send_numel = numel / nranks;
+    auto recv_buf = out->mutable_data<T>(out_dims, place);
+    auto offset = 0;
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclGroupStart());
     if (root_id == comm->rank()) {
-      PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclBcast(
-          reinterpret_cast<void*>(const_cast<T*>(x->data<T>())), numel, dtype,
+      for (auto i = 0; i < nranks; ++i) {
+        PADDLE_ENFORCE_CUDA_SUCCESS(
+            platform::dynload::ncclSend(send_buf + offset, send_numel, dtype,
                                         root_id, comm->comm(), stream));
-
-      framework::TensorCopy(*static_cast<const framework::Tensor*>(x), place,
-                            *platform::DeviceContextPool::Instance().Get(place),
-                            static_cast<framework::Tensor*>(&temp));
-    } else {
-      PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclBcast(
-          out_ptr, numel, dtype, root_id, comm->comm(), stream));
+        offset += send_numel;
       }
-
-    out_dims[0] = out_dims[0] / nranks;
-    auto start_index = out_dims[0] * comm->rank();
-    auto end_index = start_index + out_dims[0];
-    temp = temp.Slice(start_index, end_index);
-    temp.Resize(out_dims);
-    out->mutable_data<T>(out_dims, place);
-    framework::TensorCopySync(*static_cast<const framework::Tensor*>(&temp),
-                              place, static_cast<framework::Tensor*>(out));
-    out->Resize(out_dims);
+    }
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclRecv(
+        recv_buf, send_numel, dtype, root_id, comm->comm(), stream));
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclGroupEnd());
 #else
     PADDLE_ENFORCE_EQ(
         true, false,

paddle/fluid/operators/collective/c_send_op.cc

@@ -21,7 +21,19 @@ class CSendOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
-  void InferShape(framework::InferShapeContext* ctx) const override {}
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "CSend");
+    int peer = ctx->Attrs().Get<int>("peer");
+    int ring_id = ctx->Attrs().Get<int>("ring_id");
+    PADDLE_ENFORCE_GE(
+        peer, 0,
+        platform::errors::InvalidArgument(
+            "The peer (%d) for c_send_op must be non-negative.", peer));
+    PADDLE_ENFORCE_GE(
+        ring_id, 0,
+        platform::errors::InvalidArgument(
+            "The ring_id (%d) for c_send_op must be non-negative.", ring_id));
+  }
 
  protected:
   framework::OpKernelType GetExpectedKernelType(

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

@@ -44,6 +44,12 @@ class CSendOpCUDAKernel : public framework::OpKernel<T> {
     }
 
     int peer = ctx.Attr<int>("peer");
+    PADDLE_ENFORCE_LT(
+        peer, comm->nranks(),
+        platform::errors::InvalidArgument("The value of peer (%d) you set must "
+                                          "be less than comm->nranks (%d).",
+                                          peer, comm->nranks()));
+
     PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclSend(
         x->data<T>(), numel, dtype, peer, comm->comm(), stream));
     VLOG(3) << "rank " << comm->rank() << " send "

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -58,6 +58,8 @@ if(NOT WITH_GPU OR WIN32)
     LIST(REMOVE_ITEM TEST_OPS test_broadcast)
     LIST(REMOVE_ITEM TEST_OPS test_collective_reduce)
     LIST(REMOVE_ITEM TEST_OPS test_collective_sendrecv)
+    LIST(REMOVE_ITEM TEST_OPS test_collective_gather)
+    LIST(REMOVE_ITEM TEST_OPS test_collective_alltoall)
     LIST(REMOVE_ITEM TEST_OPS test_collective_scatter)
     LIST(REMOVE_ITEM TEST_OPS test_collective_reduce_api)
     LIST(REMOVE_ITEM TEST_OPS test_collective_scatter_api)

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

+# 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.
+
+from __future__ import print_function
+
+import numpy as np
+import argparse
+import os
+import sys
+import signal
+import time
+import socket
+from contextlib import closing
+from six import string_types
+import math
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.profiler as profiler
+import paddle.fluid.unique_name as nameGen
+from paddle.fluid import core
+import unittest
+from multiprocessing import Process
+import paddle.fluid.layers as layers
+from functools import reduce
+from test_collective_base import TestCollectiveRunnerBase, runtime_main
+
+
+class TestCollectiveAllToAll(TestCollectiveRunnerBase):
+    def __init__(self):
+        self.global_ring_id = 0
+
+    def get_model(self, main_prog, startup_program, rank=None):
+        ring_id = 0
+        with fluid.program_guard(main_prog, startup_program):
+            tindata = layers.data(
+                name="tindata", shape=[10, 1000], dtype='float32')
+            toutdata = layers.data(
+                name="toutdata", shape=[10, 1000], dtype='float32')
+            main_prog.global_block().append_op(
+                type="c_alltoall",
+                inputs={'X': tindata},
+                outputs={'Out': toutdata},
+                attrs={'ring_id': ring_id})
+            main_prog.global_block().append_op(
+                type="c_sync_comm_stream",
+                inputs={'X': toutdata},
+                outputs={'Out': toutdata},
+                attrs={'ring_id': ring_id})
+            return toutdata
+
+
+if __name__ == "__main__":
+    runtime_main(TestCollectiveAllToAll, "alltoall", 0)

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

+# 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.
+
+from __future__ import print_function
+
+import numpy as np
+import argparse
+import os
+import sys
+import signal
+import time
+import socket
+from contextlib import closing
+from six import string_types
+import math
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.profiler as profiler
+import paddle.fluid.unique_name as nameGen
+from paddle.fluid import core
+import unittest
+from multiprocessing import Process
+import paddle.fluid.layers as layers
+from functools import reduce
+from test_collective_base import TestCollectiveRunnerBase, runtime_main
+
+
+class TestCollectiveGather(TestCollectiveRunnerBase):
+    def __init__(self):
+        self.global_ring_id = 0
+
+    def get_model(self, main_prog, startup_program, rank=None):
+        ring_id = 0
+        with fluid.program_guard(main_prog, startup_program):
+            tindata = layers.data(
+                name="tindata", shape=[10, 1000], dtype='float32')
+            toutdata = layers.data(
+                name="toutdata", shape=[20, 1000], dtype='float32')
+            main_prog.global_block().append_op(
+                type="c_gather",
+                inputs={'X': tindata},
+                outputs={'Out': toutdata},
+                attrs={'ring_id': ring_id,
+                       'nranks': 2,
+                       'root': 1})
+            main_prog.global_block().append_op(
+                type="c_sync_comm_stream",
+                inputs={'X': toutdata},
+                outputs={'Out': toutdata},
+                attrs={'ring_id': ring_id})
+            return toutdata
+
+
+if __name__ == "__main__":
+    runtime_main(TestCollectiveGather, "gather", 0)

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

@@ -36,7 +36,7 @@ from functools import reduce
 from test_collective_base import TestCollectiveRunnerBase, runtime_main
 
 
-class TestCollectiveScatter(TestCollectiveRunnerBase):
+class TestCollectiveSendRecv(TestCollectiveRunnerBase):
     def __init__(self):
         self.global_ring_id = 0
 
@@ -48,7 +48,7 @@ class TestCollectiveScatter(TestCollectiveRunnerBase):
             if rank == 0:
                 main_prog.global_block().append_op(
                     type="c_recv",
-                    inputs={'Out': tindata},
+                    outputs={'Out': tindata},
                     attrs={'ring_id': ring_id,
                            'peer': 1})
             else:
@@ -59,11 +59,11 @@ class TestCollectiveScatter(TestCollectiveRunnerBase):
                            'peer': 0})
             main_prog.global_block().append_op(
                 type="c_sync_comm_stream",
-                inputs={'X': toutdata},
-                outputs={'Out': toutdata},
+                inputs={'X': tindata},
+                outputs={'Out': tindata},
                 attrs={'ring_id': ring_id})
             return tindata
 
 
 if __name__ == "__main__":
-    runtime_main(TestCollectiveScatter, "scatter", 0)
+    runtime_main(TestCollectiveSendRecv, "sendrecv", 0)

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

+#   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.
+
+from __future__ import print_function
+import unittest
+import numpy as np
+
+from test_collective_base import TestDistBase
+
+
+class TestCAllToAllOp(TestDistBase):
+    def _setup_config(self):
+        pass
+
+    def test_alltoall(self):
+        self.check_with_place("collective_alltoall_op.py", "alltoall")
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -261,6 +261,22 @@ class TestDistBase(unittest.TestCase):
         elif col_type == "sendrecv":
             need_result = input2
             self.assertTrue(np.allclose(tr0_out, need_result))
+        elif col_type == "gather":
+            need_result = np.vstack((input1, input2))
+            self.assertTrue(np.allclose(tr1_out, need_result))
+        elif col_type == "alltoall":
+            temp11, temp12 = np.split(input1, 2)
+            temp21, temp22 = np.split(input2, 2)
+            need_result1 = np.hstack((temp11, temp21))
+            need_result2 = np.hstack((temp12, temp22))
+            print("input1:", input1)
+            print("input2:", input2)
+            print("need_result1:", need_result1)
+            print("need_result2:", need_result2)
+            print("tr0_out:", tr0_out)
+            print("tr1_out:", tr1_out)
+            self.assertTrue(np.allclose(tr1_out, need_result1))
+            self.assertTrue(np.allclose(tr2_out, need_result2))
         elif col_type == "reduce_scatter":
             tmp = input1 + input2
             need_result1 = tmp[0:tmp.shape[0] // 2]

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

+#   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.
+
+from __future__ import print_function
+import unittest
+import numpy as np
+
+from test_collective_base import TestDistBase
+
+
+class TestCGatherOp(TestDistBase):
+    def _setup_config(self):
+        pass
+
+    def test_gather(self):
+        self.check_with_place("collective_gather_op.py", "gather")
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -19,7 +19,7 @@ import numpy as np
 from test_collective_base import TestDistBase
 
 
-class TestCScatterOp(TestDistBase):
+class TestCSendRecvOp(TestDistBase):
     def _setup_config(self):
         pass