graphengine/ge/graph/passes/mds_pass.cc

/**
 * Copyright 2021 Huawei Technologies Co., Ltd
 *
 * 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 "./mds_pass.h"

namespace ge {
Status ModelDeploySchedulerPass::Run(ComputeGraphPtr graph) {
  GE_CHECK_NOTNULL(graph);
  compute_graph_ = graph;
  if (!MdsUtils::IsMDSNeeded()) {
    return SUCCESS;
  }
  GELOGI("[MDS][%s] start to deploy.", GetGraphName());
  MDS_REQUIRE_SUCCESS(SMDPProcess(), "[MDS][SMDPProcess] failed, graph_name:[%s]", GetGraphName());
  MDS_REQUIRE_SUCCESS(CutProcess(), "[MDS][CutProcess] failed, graph_name:[%s]", GetGraphName());
  MDS_REQUIRE_SUCCESS(SMDPProcess(false), "[MDS][SMDPProcess] failed, graph_name:[%s]", GetGraphName());
  MDS_REQUIRE_SUCCESS(SwapProcess(), "[MDS][SwapProcess] failed, graph_name:[%s]", GetGraphName());
  MDS_REQUIRE_SUCCESS(PiplineProcess(), "[MDS][PiplineProcess] failed, graph_name:[%s]", GetGraphName());
  MDS_REQUIRE_SUCCESS(SetDeployInfo(), "[MDS][SetDeployInfo] failed, graph_name:[%s]", GetGraphName());

  GELOGI("[MDS][%s] deploy successfully.", graph->GetName().c_str());
  return SUCCESS;
}

Status ModelDeploySchedulerPass::CutProcess() {
  GE_CHECK_NOTNULL(compute_graph_);
  if (!compute_graph_->GetAllSubgraphs().empty() || compute_graph_->GetParentGraph() != nullptr) {
    GELOGW("[MDS][CutProcess] graph with subgraphs is not supported now. graph_name:[%s]", GetGraphName());
    return SUCCESS;
  }
  auto type = MdsUtils::TryGetGraphCutType(compute_graph_);
  switch (type) {
    case kCutN:MDS_REQUIRE_SUCCESS(CutNProcessImply(compute_graph_),
                                   "[MDS][CutNProcessImply] failed, graph_name:[%s]",
                                   GetGraphName());
      break;
    case kCutH:MDS_REQUIRE_SUCCESS(CutHProcessImply(compute_graph_),
                                   "[MDS][CutHProcessImply] failed, graph_name:[%s]",
                                   GetGraphName());
      break;
    case kDynamicCutN:MDS_REQUIRE_SUCCESS(CutNProcessImply(compute_graph_, true),
                                          "[MDS][CutNProcessImply] failed, graph_name:[%s]",
                                          GetGraphName());
      break;
    case kDynamicCutH:MDS_REQUIRE_SUCCESS(CutHProcessImply(compute_graph_, true),
                                          "[MDS][CutHProcessImply] failed, graph_name:[%s]",
                                          GetGraphName());
      break;
    case kDynamicCutAll:MDS_REQUIRE_SUCCESS(DynamicCutAll(compute_graph_),
                                            "[MDS][DynamicCutAll] failed, graph_name:[%s]",
                                            GetGraphName());
      break;
    default:GELOGI("[MDS][CutProcess] could not cut, just return. graph_name:[%s]", GetGraphName());
      return SUCCESS;
  }

}

Status ModelDeploySchedulerPass::CutNProcessImply(const ComputeGraphPtr &compute_graph, bool is_dynamic) {
  GE_CHECK_NOTNULL(compute_graph);
  // step 0: Cut
  for (const auto &node : compute_graph->GetDirectNode()) {
    auto op_kernel = mds_cut_pass::GetKernelByType(node);
    if (op_kernel == nullptr) {
      op_kernel = DeploySchedulerKernel::Instance();
    }
    if (is_dynamic) {
      MDS_REQUIRE_SUCCESS(op_kernel->DynamicCutN(node),
                          "[MDS][DYNAMIC_CUTN] failed, node:[%s]",
                          node->GetName().c_str());
    } else {
      MDS_REQUIRE_SUCCESS(op_kernel->CutN(node), "[MDS][CUTN] failed, node:[%s]", node->GetName().c_str());
    }
    bool is_grad_compute_node = false;
    if (ge::AttrUtils::GetBool(node->GetOpDesc(), ATTR_NAME_GRADIENT_NODE, is_grad_compute_node)
        && is_grad_compute_node) {
      grad_compute_nodes_.push_back(node);
    }
  }
  //TODO:针对单输出多引用插入的allgather，allreduce节点做广度融合优化
  MDS_REQUIRE_SUCCESS(HcomNodeFusionProcess(),
                      "[MDS][CUTN][HcomNodeFusionProcess] failed, compute graph:[%s]",
                      compute_graph->GetName().c_str());
  return SUCCESS;
}

Status ModelDeploySchedulerPass::CutHProcessImply(const ComputeGraphPtr &compute_graph, bool is_dynamic) {
  GE_CHECK_NOTNULL(compute_graph);
  for (NodePtr &node : compute_graph->GetDirectNode()) {
    auto op_kernel = mds_cut_pass::GetKernelByType(node);
    if (op_kernel == nullptr) {
      op_kernel = DeploySchedulerKernel::Instance();
    }
    if (is_dynamic) {
      MDS_REQUIRE_SUCCESS(op_kernel->DynamicCutH(node),
                          "[MDS][DYNAMIC_CUTH] failed, node:[%s]",
                          node->GetName().c_str());
    } else {
      MDS_REQUIRE_SUCCESS(op_kernel->CutH(node), "[MDS][CUTH] failed, node:[%s]", node->GetName().c_str());

    }
  }
  return SUCCESS;
}

Status ModelDeploySchedulerPass::DynamicCutAll(const ComputeGraphPtr &compute_graph) {
  std::vector<NodePtr> input_nodes;
  std::vector<NodePtr> output_nodes;
  auto compute_graph0 = GraphUtils::CloneGraph(compute_graph, "", input_nodes, output_nodes);
  auto compute_graph1 = GraphUtils::CloneGraph(compute_graph, "", input_nodes, output_nodes);
  MDS_REQUIRE_SUCCESS(CutNProcessImply(compute_graph0, true),
                      "[MDS][CutNProcessImply] failed, graph_name:[%s]",
                      compute_graph0->GetName().c_str());
  MDS_REQUIRE_SUCCESS(CutHProcessImply(compute_graph1, true),
                      "[MDS][CutHProcessImply] failed, graph_name:[%s]",
                      compute_graph1->GetName().c_str());
//TODO:创建case节点，把两个图放在case的两个分支下，case节点添加到原来的compute_graph中,构造case节点的输入
  return SUCCESS;
}

Status ModelDeploySchedulerPass::SMDPProcess(bool before_cut) {
  if (before_cut) {
    MDS_REQUIRE_SUCCESS(SMDPModelState(), "[SMDPProcess][SMDPModelState] failed, graph_name:[%s]", GetGraphName());
    MDS_REQUIRE_SUCCESS(SMDPWeight(), "[SMDPProcess][SMDPWeight] failed, graph_name:[%s]", GetGraphName());
  } else {
    MDS_REQUIRE_SUCCESS(SMDPGradient(), "[SMDPProcess][SMDPGradient] failed, graph_name:[%s]", GetGraphName());
  }
  return SUCCESS;
}

Status ModelDeploySchedulerPass::SetDeployInfo() {
  vector<GeAttrValue::NAMED_ATTRS> deployInfo;
  REQUIRE (!ge::AttrUtils::GetListNamedAttrs(compute_graph_, ATTR_NAME_DEPLOY_INFO, deployInfo),
           "%s already has deployed before!",
           GetGraphName());
  std::multimap<DeviceId, GraphInputs> deploys;
  for (int64_t j = 0; j < kDeployNumber; j++) {
    int64_t device_id = j;
    GraphInputs graph_inputs;
    // For now, only one input_node in input_nodes
    for (const auto &input_node : MdsUtils::GetInputNodes()) {
      GE_CHECK_NOTNULL(input_node);
      GeTensorPtr graph_input = MakeShared<GeTensor>(input_node->GetOpDesc()->GetOutputDesc(0));
      vector<uint8_t> data{static_cast<uint8_t>(device_id)};
      graph_input->SetData(data);
      graph_inputs.push_back(graph_input);
    }
    deploys.emplace(j, graph_inputs);
  }
  return MdsUtils::SetDeployInfo(compute_graph_, deploys);
}

Status ModelDeploySchedulerPass::SwapProcess() {
  return SUCCESS;
}
Status ModelDeploySchedulerPass::PiplineProcess() {
  return SUCCESS;
}
Status ModelDeploySchedulerPass::HcomNodeFusionProcess() {
  return SUCCESS;
}
Status ModelDeploySchedulerPass::SMDPModelState() {
  return SUCCESS;
}
Status ModelDeploySchedulerPass::SMDPWeight() {
  return SUCCESS;
}
Status ModelDeploySchedulerPass::SMDPGradient() {
  //TDOD:标识buffer poolid
  return SUCCESS;
}
}