mindspore/mindspore/ccsrc/backend/session/gpu_session.cc

/**
 * Copyright 2019-2020 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 "debug/anf_ir_utils.h"
#include "backend/session/gpu_session.h"
#include "runtime/device/gpu/kernel_info_setter.h"
#include "runtime/device/gpu/gpu_kernel_build.h"
#include "runtime/device/gpu/gpu_kernel_runtime.h"
#include "runtime/device/gpu/gpu_stream_assign.h"
#include "backend/optimizer/common/optimizer.h"
#include "backend/optimizer/common/pass_manager.h"
#include "backend/optimizer/common/helper.h"
#include "backend/optimizer/pass/communication_op_fusion.h"
#include "backend/optimizer/pass/getitem_tuple.h"
#include "backend/optimizer/gpu/adam_weight_decay_fusion.h"
#include "backend/optimizer/gpu/adam_fusion.h"
#include "backend/optimizer/gpu/replace_bn_cast_fusion.h"
#include "backend/optimizer/gpu/replace_bn_grad_cast_fusion.h"
#include "backend/optimizer/gpu/replace_bn_grad_cast2_fusion.h"
#include "backend/optimizer/gpu/replace_momentum_cast_fusion.h"
#include "backend/optimizer/gpu/replace_addn_fusion.h"
#include "runtime/device/kernel_runtime_manager.h"
#include "predict/predict.h"
#include "common/utils.h"
#include "common/trans.h"
#include "utils/context/ms_context.h"
#include "utils/base_ref_extends.h"
#include "debug/tensor_load.h"

namespace mindspore {
namespace session {
namespace gpu {
using AnfAlgo = mindspore::session::AnfRuntimeAlgorithm;

void GPUSession::SelectKernel(const std::shared_ptr<KernelGraph> &kernel_graph) const {
  MS_EXCEPTION_IF_NULL(kernel_graph);
  for (const auto &kernel_node : kernel_graph->execution_order()) {
    MS_EXCEPTION_IF_NULL(kernel_node);
    device::gpu::SetKernelInfo(kernel_node);
  }
}

void GPUSession::StartKernelRT() const {
  auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
  MS_EXCEPTION_IF_NULL(runtime_instance);
  if (!runtime_instance->Init()) {
    MS_LOG(EXCEPTION) << "GPU start kernel runtime failed";
  }
}

void GPUSession::Optimize(const std::shared_ptr<KernelGraph> &kernel_graph) {
  MS_EXCEPTION_IF_NULL(kernel_graph);
  auto optimizer = std::make_shared<opt::GraphOptimizer>();
  auto pm = std::make_shared<opt::PassManager>();
  pm->AddPass(std::make_shared<opt::AdamWeightDecayFusion>());
  pm->AddPass(std::make_shared<opt::AdamFusion>());
  pm->AddPass(std::make_shared<opt::ReplaceBNCastFusion>());
  pm->AddPass(std::make_shared<opt::ReplaceBNGradCastFusion>());
  pm->AddPass(std::make_shared<opt::ReplaceBNGradCast2Fusion>());
  pm->AddPass(std::make_shared<opt::ReplaceMomentumCastFusion>());
  pm->AddPass(std::make_shared<opt::ReplaceAddNFusion>());
  optimizer->AddPassManager(pm);
  (void)optimizer->Optimize(kernel_graph);
  kernel_graph->SetExecOrderByDefault();
}

void GPUSession::HardwareOptimize(const std::shared_ptr<KernelGraph> &kernel_graph) {
  auto optimizer = std::make_shared<opt::GraphOptimizer>();
  auto pm = std::make_shared<opt::PassManager>();
  pm->AddPass(std::make_shared<opt::AllReduceFusion>());
  pm->AddPass(std::make_shared<opt::GetitemTuple>());
  optimizer->AddPassManager(pm);
  (void)optimizer->Optimize(kernel_graph);
  kernel_graph->SetExecOrderByDefault();
}

void GPUSession::AssignStream(const std::shared_ptr<KernelGraph> &kernel_graph) {
  MS_EXCEPTION_IF_NULL(kernel_graph);
  device::gpu::AssignGpuStream(kernel_graph);
}

void GPUSession::BuildKernel(const std::shared_ptr<KernelGraph> &kernel_graph) const {
  device::gpu::GpuBuild(kernel_graph);
}

void GPUSession::AllocateMemory(KernelGraph *kernel_graph) const {
  MS_EXCEPTION_IF_NULL(kernel_graph);
  auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
  MS_EXCEPTION_IF_NULL(runtime_instance);
  runtime_instance->AssignMemory(kernel_graph);
}

void GPUSession::RunOpAllocateMemory(const ValuePtr &pre_output_value,
                                     const std::vector<tensor::TensorPtr> &input_tensors,
                                     KernelGraph *kernel_graph) const {
  MS_EXCEPTION_IF_NULL(kernel_graph);
  auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
  MS_EXCEPTION_IF_NULL(runtime_instance);
  runtime_instance->RunOpAssignMemory(pre_output_value, input_tensors, kernel_graph);
}

void GPUSession::RunOpClearMemory(KernelGraph *kernel_graph) const {
  MS_EXCEPTION_IF_NULL(kernel_graph);
  auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
  MS_EXCEPTION_IF_NULL(runtime_instance);
  runtime_instance->RunOpClearMemory(kernel_graph);
}

void GPUSession::LoadInputData(const std::shared_ptr<KernelGraph> &kernel_graph,
                               const std::vector<tensor::TensorPtr> &inputs_const) const {
  std::vector<tensor::TensorPtr> inputs(inputs_const);
  MS_EXCEPTION_IF_NULL(kernel_graph);
  auto input_nodes = kernel_graph->inputs();
  auto ms_context = MsContext::GetInstance();
  MS_EXCEPTION_IF_NULL(ms_context);

  for (size_t i = 0; i < inputs.size(); ++i) {
    auto tensor = inputs[i];
    MS_EXCEPTION_IF_NULL(tensor);
    auto input_node = input_nodes[i];
    MS_EXCEPTION_IF_NULL(input_node);
    if (input_node->isa<Parameter>() && AnfAlgo::OutputAddrExist(input_node, 0)) {
      auto pk_node = input_node->cast<ParameterPtr>();
      auto device_address = AnfAlgo::GetMutableOutputAddr(pk_node, 0);
      auto tensor_address = std::dynamic_pointer_cast<device::DeviceAddress>(tensor->device_address());
      bool need_sync = false;
      if (ms_context->enable_pynative_infer()) {
        if (tensor_address == nullptr || tensor_address != device_address) {
          need_sync = true;
        }
      } else if (tensor->is_dirty() || tensor_address == nullptr) {
        need_sync = true;
      } else if (tensor_address != device_address) {
        if (tensor_address->DeviceType() == device_address->DeviceType()) {
          AnfAlgo::SetOutputAddr(tensor_address, 0, pk_node.get());
        } else {
          need_sync = true;
        }
      }
      if (need_sync) {
        tensor->set_device_address(device_address);
        MS_EXCEPTION_IF_NULL(device_address);
        if (!device_address->SyncHostToDevice(trans::GetRuntimePaddingShape(pk_node, 0),
                                              LongToSize(tensor->data().nbytes()), tensor->data_type(),
                                              tensor->data_c())) {
          MS_LOG(EXCEPTION) << "SyncHostToDevice failed.";
        }
      }
    }
    tensor->set_dirty(false);
  }
}

void GPUSession::Execute(const std::shared_ptr<KernelGraph> &kernel_graph) const {
  auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
  MS_EXCEPTION_IF_NULL(runtime_instance);
#ifdef ENABLE_DEBUGGER
  if (!runtime_instance->Run(kernel_graph.get(), debugger_.get())) {
#else
  if (!runtime_instance->Run(kernel_graph.get())) {
#endif
    MS_LOG(EXCEPTION) << "GPU execute graph failed!";
  }
}

GraphId GPUSession::CompileGraph(const AnfNodePtrList &lst, const AnfNodePtrList &outputs) {
  // Construct graph, if successfully, graph_sum_ + 1
  auto graph_id = graph_sum_;
  auto graph = ConstructKernelGraph(lst, outputs);
  MS_EXCEPTION_IF_NULL(graph);
  // Prepare ms context info for dump .pb graph
  auto context_ptr = MsContext::GetInstance();
  MS_EXCEPTION_IF_NULL(context_ptr);
  bool save_graphs = context_ptr->save_graphs_flag();
  // Optimize
  Optimize(graph);
  // Select kernel build info
  SelectKernel(graph);
#if (ENABLE_CPU && (ENABLE_D || ENABLE_GPU))
  // Assign parameter keys.
  AssignParamKey(graph);
#endif
  // Convert kernel Graph to model
  predictmodel::StepConvertGraph(graph);
  // Start gpu kernel runtime
  StartKernelRT();
  // Dump .pb graph before hardware optimization
  if (save_graphs) {
    DumpIRProto(graph, "before_hwopt_" + std::to_string(graph_id));
  }
  // HardwareOptimize
  HardwareOptimize(graph);
  // Dump .pb graph after hardware optimization
  if (save_graphs) {
    DumpIRProto(graph, "after_hwopt_" + std::to_string(graph_id));
  }
  // Assign CUDA streams
  AssignStream(graph);
  // Hide NoOp from execution graph
  opt::HideNopNode(graph.get());
  // Build kernel if node is cnode
  BuildKernel(graph);
  // Set graph execution order before memory alloc, ensure that memory alloc is according to the reorder graph
  auto execution_order = graph->execution_order();
  Reorder(&execution_order);
  graph->set_execution_order(execution_order);
  // Get summary nodes.
  SetSummaryNodes(graph.get());
  // Remove NoOp from execution graph
  opt::RemoveNopNode(graph.get());
  // Set graph manager.
  MS_EXCEPTION_IF_NULL(context_);
  FuncGraphManagerPtr manager = MakeManager({graph});
  context_->AddManager(manager);
  if (manager) {
    manager->AddFuncGraph(graph);
    graph->set_manager(manager);
  }
  // Alloc memory, including static memory and dynamic memory
  AllocateMemory(graph.get());
  return graph_id;
}

void GPUSession::RunGraph(const GraphId &graph_id, const std::vector<tensor::TensorPtr> &inputs, VectorRef *outputs) {
  auto &kernel_graph = graphs_[graph_id];
#ifdef ENABLE_DEBUGGER
  PreIterationDbg(kernel_graph);
#endif
  // Load input data from user input
  LoadInputData(kernel_graph, inputs);
#if (ENABLE_CPU && (ENABLE_D || ENABLE_GPU))
  // Initialize parameter server
  if (!ps_init_) {
    InitPSParamAndOptim(kernel_graph, inputs);
  }
#endif
  MS_EXCEPTION_IF_NULL(kernel_graph);
  // Convert inputs to model
  predictmodel::StepConvertWeight(inputs);
  {
    py::gil_scoped_release gil_release;
    // Run graph on GPU
    Execute(kernel_graph);
  }
#ifdef ENABLE_DEBUGGER
  PostLoadTensor(kernel_graph);
#endif
  // Get result from GPU
  UpdateOutputs(kernel_graph, outputs, inputs);
  // Summary
  auto context_ptr = MsContext::GetInstance();
  MS_EXCEPTION_IF_NULL(context_ptr);
  if (context_ptr->enable_gpu_summary()) {
    Summary(kernel_graph.get());
  }
#ifdef ENABLE_DEBUGGER
  PostIterationDbg(kernel_graph);
#endif
}

void GPUSession::BuildOp(const OpRunInfo &op_run_info, const GraphInfo &graph_info,
                         const std::vector<tensor::TensorPtr> &input_tensors, const std::vector<int> &tensors_mask) {
  // Check if the graph cache exists.
  if (run_op_graphs_.find(graph_info) != run_op_graphs_.end()) {
    return;
  }
  // Prepare the graph
  auto kernel_graph = ConstructSingleOpGraph(op_run_info, input_tensors, tensors_mask);
  MS_EXCEPTION_IF_NULL(kernel_graph);
  SelectKernel(kernel_graph);
  StartKernelRT();
  // Hide NoOp from execution graph
  opt::HideNopNode(kernel_graph.get());
  BuildKernel(kernel_graph);
  run_op_graphs_[graph_info] = kernel_graph;
}

py::tuple GPUSession::RunOp(const OpRunInfo &op_run_info, const GraphInfo &graph_info,
                            const std::vector<tensor::TensorPtr> &input_tensors) {
  auto kernel_graph = run_op_graphs_[graph_info];
  MS_EXCEPTION_IF_NULL(kernel_graph);
  // Remove NoOp from execution graph
  opt::RemoveNopNode(kernel_graph.get());
  RunOpAllocateMemory(op_run_info.value, input_tensors, kernel_graph.get());
  // Execute the computation
  LoadInputData(kernel_graph, input_tensors);
  Execute(kernel_graph);
  // Fetch outputs
  VectorRef outputs;
  UpdateOutputs(kernel_graph, &outputs, input_tensors);
  // Trans output to tuple
  auto output_tensors = TransformBaseRefListToTuple(outputs);
  if (!utils::isa<PyObjectRef>(output_tensors) ||
      !py::isinstance<py::tuple>(utils::cast<PyObjectRef>(output_tensors).object_)) {
    MS_EXCEPTION(NotSupportError) << "The output tensors should be a tuple !";
  }
  py::object tuple_obj = utils::cast<PyObjectRef>(output_tensors).object_;
  py::tuple tuple_tensors = py::cast<py::tuple>(tuple_obj);
  RunOpClearMemory(kernel_graph.get());
  return tuple_tensors;
}

#ifdef ENABLE_DEBUGGER
void GPUSession::Dump(const std::shared_ptr<KernelGraph> &kernel_graph) const {
#ifdef ENABLE_DUMP_E2E
  MS_EXCEPTION_IF_NULL(kernel_graph);
  auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
  MS_EXCEPTION_IF_NULL(runtime_instance);
  (void)runtime_instance->DumpData(kernel_graph.get(), debugger_.get());
#endif
}

bool GPUSession::DumpDataEnabledIteration() const {
  auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
  MS_EXCEPTION_IF_NULL(runtime_instance);
  return runtime_instance->DumpDataEnabledIteration();
}

void GPUSession::PreIterationDbg(const std::shared_ptr<KernelGraph> &kernel_graph) const {
  if (debugger_) {
    debugger_->PreExecute(kernel_graph);
  }
  PreLoadTensor(kernel_graph);
}

void GPUSession::PostIterationDbg(const std::shared_ptr<KernelGraph> &kernel_graph) const {
  bool dump_enabled = DumpDataEnabledIteration();
  // debug used for dump
  if (debugger_ && dump_enabled) {
    Dump(kernel_graph);
  }
  if (debugger_) {
    debugger_->PostExecute();
  }
}

void GPUSession::PreLoadTensor(const std::shared_ptr<KernelGraph> &kernel_graph) const {
  bool dump_enabled = DumpDataEnabledIteration();
  if (!(debugger_ && (debugger_->debugger_enabled() || dump_enabled))) {
    return;
  }
  MS_EXCEPTION_IF_NULL(kernel_graph);
  auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
  MS_EXCEPTION_IF_NULL(runtime_instance);
  DebugServices *debug_services = debugger_->debug_services();
  TensorLoader *tensor_loader = debug_services->tensor_loader();
  tensor_loader->EmptyTensor();
  uint32_t iter_num = tensor_loader->GetIterNum();
  tensor_loader->set_iter_num(++iter_num);
}

void GPUSession::PostLoadTensor(const std::shared_ptr<KernelGraph> &kernel_graph) const {
  bool dump_enabled = DumpDataEnabledIteration();
  if (!(debugger_ && (debugger_->debugger_enabled() || dump_enabled))) {
    return;
  }
  MS_EXCEPTION_IF_NULL(kernel_graph);
  auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
  MS_EXCEPTION_IF_NULL(runtime_instance);
  DebugServices *debug_services = debugger_->debug_services();
  TensorLoader *tensor_loader = debug_services->tensor_loader();
  tensor_loader->EmptyPrevTensor();
}
#endif

}  // namespace gpu
}  // namespace session
}  // namespace mindspore