/* 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" #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) #include "paddle/fluid/platform/collective_helper.h" #include "paddle/fluid/platform/device/gpu/nccl_helper.h" #endif #include "paddle/fluid/distributed/collective/process_group.h" #include "paddle/fluid/framework/convert_utils.h" #include "paddle/phi/api/include/tensor.h" namespace paddle { namespace operators { template class CAllGatherOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) auto in = ctx.Input("X"); auto out = ctx.Output("Out"); ncclDataType_t dtype = platform::ToNCCLDataType(framework::TransToProtoVarType(in->dtype())); int nranks = ctx.Attr("nranks"); int rid = ctx.Attr("ring_id"); auto map = distributed::ProcessGroupMapFromGid::getInstance(); if (map->has(rid)) { // Use ProcessGroup distributed::ProcessGroup* pg = map->get(rid); std::vector in_tensor; std::vector out_tensor; in_tensor.push_back(*in); out_tensor.push_back(*out); auto task = pg->AllGather(in_tensor, out_tensor); task->Wait(); return; } auto place = ctx.GetPlace(); auto comm = platform::NCCLCommContext::Instance().Get(rid, place); PADDLE_ENFORCE_EQ( nranks, comm->nranks(), platform::errors::InvalidArgument( "nranks: %s should equal to %s", nranks, comm->nranks())); int64_t send_numel = in->numel(); const T* send_buff = in->data(); T* recv_buff = out->mutable_data(place); gpuStream_t stream = nullptr; if (ctx.Attr("use_calc_stream")) { // should ExecutionContext for calc stream. stream = ctx.cuda_device_context().stream(); } else { stream = comm->stream(); } PADDLE_ENFORCE_GPU_SUCCESS( platform::dynload::ncclAllGather(send_buff, recv_buff, send_numel, static_cast(dtype), comm->comm(), stream)); #else PADDLE_THROW(platform::errors::PreconditionNotMet( "PaddlePaddle should compile with GPU.")); #endif } }; } // namespace operators } // namespace paddle namespace ops = paddle::operators; namespace plat = paddle::platform; PD_REGISTER_STRUCT_KERNEL(c_allgather, GPU, ALL_LAYOUT, ops::CAllGatherOpCUDAKernel, float, double, #if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000 plat::bfloat16, #endif int, uint8_t, int8_t, int64_t, bool, plat::float16) { }