c_allgather_op.cu.cc 3.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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_allgather_op.h"

17
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
18
#include "paddle/fluid/platform/collective_helper.h"
19
#include "paddle/fluid/platform/device/gpu/nccl_helper.h"
20
#endif
W
Wen Sun 已提交
21
#include "paddle/fluid/distributed/collective/process_group.h"
22
#include "paddle/fluid/framework/convert_utils.h"
23
#include "paddle/phi/api/include/tensor.h"
24 25 26 27 28 29 30 31

namespace paddle {
namespace operators {

template <typename T>
class CAllGatherOpCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
32
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
33 34
    auto in = ctx.Input<phi::DenseTensor>("X");
    auto out = ctx.Output<phi::DenseTensor>("Out");
35 36
    ncclDataType_t dtype =
        platform::ToNCCLDataType(framework::TransToProtoVarType(in->dtype()));
37 38 39

    int nranks = ctx.Attr<int>("nranks");
    int rid = ctx.Attr<int>("ring_id");
40 41 42 43 44 45 46 47 48 49 50 51
    auto map = distributed::ProcessGroupMapFromGid::getInstance();
    if (map->has(rid)) {
      // Use ProcessGroup
      distributed::ProcessGroup* pg = map->get(rid);
      std::vector<phi::DenseTensor> in_tensor;
      std::vector<phi::DenseTensor> out_tensor;
      in_tensor.push_back(*in);
      out_tensor.push_back(*out);
      auto task = pg->AllGather(in_tensor, out_tensor);
      task->Wait();
      return;
    }
52 53
    auto place = ctx.GetPlace();
    auto comm = platform::NCCLCommContext::Instance().Get(rid, place);
M
MRXLT 已提交
54
    PADDLE_ENFORCE_EQ(
55 56 57 58
        nranks,
        comm->nranks(),
        platform::errors::InvalidArgument(
            "nranks: %s should equal to %s", nranks, comm->nranks()));
59 60 61 62 63 64 65 66 67

    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>();

68
    gpuStream_t stream = nullptr;
69
    if (ctx.Attr<bool>("use_calc_stream")) {
70 71
      // should ExecutionContext for calc stream.
      stream = ctx.cuda_device_context().stream();
72 73 74 75
    } else {
      stream = comm->stream();
    }

76 77 78 79 80 81 82
    PADDLE_ENFORCE_GPU_SUCCESS(
        platform::dynload::ncclAllGather(send_buff,
                                         recv_buff,
                                         send_numel,
                                         static_cast<ncclDataType_t>(dtype),
                                         comm->comm(),
                                         stream));
83
#else
M
MRXLT 已提交
84 85
    PADDLE_THROW(platform::errors::PreconditionNotMet(
        "PaddlePaddle should compile with GPU."));
86 87 88 89 90 91 92
#endif
  }
};

}  // namespace operators
}  // namespace paddle

93 94 95
namespace ops = paddle::operators;
namespace plat = paddle::platform;

96 97
REGISTER_OP_CUDA_KERNEL(c_allgather,
                        ops::CAllGatherOpCUDAKernel<float>,
98
                        ops::CAllGatherOpCUDAKernel<double>,
L
LiYuRio 已提交
99
#if NCCL_VERSION_CODE >= 21000
100 101
                        ops::CAllGatherOpCUDAKernel<plat::bfloat16>,
#endif
102
                        ops::CAllGatherOpCUDAKernel<int>,
103 104
                        ops::CAllGatherOpCUDAKernel<uint8_t>,
                        ops::CAllGatherOpCUDAKernel<int8_t>,
105
                        ops::CAllGatherOpCUDAKernel<int64_t>,
106
                        ops::CAllGatherOpCUDAKernel<bool>,
107
                        ops::CAllGatherOpCUDAKernel<plat::float16>);