Pre Merge pull request !588 from 储星/single_op

ge/single_op/single_op.cc

@@ -25,6 +25,7 @@
 #include "graph/load/new_model_manager/model_utils.h"
 #include "runtime/mem.h"
 #include "single_op/single_op_manager.h"
+#include "single_op/task/build_task_utils.h"
 #include "graph/load/new_model_manager/model_manager.h"
 
 namespace ge {
@@ -77,7 +78,8 @@ Status ProfilingTaskInfo(OpTask *op_task) {
 }
 }  // namespace
 
-SingleOp::SingleOp(std::mutex *stream_mutex, rtStream_t stream) : stream_mutex_(stream_mutex), stream_(stream) {
+SingleOp::SingleOp(StreamResource *stream_resource, std::mutex *stream_mutex, rtStream_t stream)
+    : stream_resource_(stream_resource), stream_mutex_(stream_mutex), stream_(stream) {
 }
 
 FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY SingleOp::~SingleOp() {
@@ -159,37 +161,6 @@ Status SingleOp::UpdateArgs(const std::vector<DataBuffer> &inputs, const std::ve
       *arg_addr = args_[i];
     }
   }
-  // update aicpu_TF or aicpu_CC args
-  for (auto &task : tasks_) {
-    size_t io_addr_num = args_.size();
-    if (task->GetOpTaskType() == OP_TASK_AICPU) {
-      GELOGD("Update aicpu_TF task args");
-      task->SetIoAddrsForDump(args_);
-      auto *dst_io_addr = const_cast<uintptr_t *>(reinterpret_cast<const uintptr_t *>(task->GetIOAddr()));
-      GE_CHECK_NOTNULL(dst_io_addr);
-      auto rt_ret = rtMemcpyAsync(dst_io_addr,
-                                  sizeof(uint64_t) * args_.size(),
-                                  &args_[0],
-                                  sizeof(uint64_t) * args_.size(),
-                                  RT_MEMCPY_HOST_TO_DEVICE_EX,
-                                  stream_);
-      if (rt_ret != RT_ERROR_NONE) {
-        GELOGE(rt_ret, "rtMemcpyAsync addresses failed, ret = %d", rt_ret);
-        return rt_ret;
-      }
-    } else if (task->GetOpTaskType() == OP_TASK_AICPUCC) {
-      GELOGD("Update aicpu_CC task args");
-      const uintptr_t *task_io_addr = reinterpret_cast<const uintptr_t *>(task->GetIOAddr());
-      GE_CHECK_NOTNULL(task_io_addr);
-      auto io_addr = reinterpret_cast<uint64_t *>(const_cast<uintptr_t *>(task_io_addr));
-      for (size_t i = 0; i < io_addr_num; ++i) {
-        io_addr[i] = static_cast<uintptr_t>(args_[i]);
-      }
-    } else {
-      GELOGW("Only TF_kernel aicpu and aicpu_CC are supported, but got %u", task->GetOpTaskType());
-      continue;
-    }
-  }
   return SUCCESS;
 }
 
@@ -200,7 +171,19 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status SingleOp::ExecuteAsync(c
     return ret;
   }
 
+  GE_CHECK_NOTNULL(stream_resource_);
   std::lock_guard<std::mutex> lk(*stream_mutex_);
+  auto current_mem_base = stream_resource_->GetMemoryBase();
+  if (running_param_->mem_base != current_mem_base) {
+    running_param_->mem_base = const_cast<uint8_t *>(current_mem_base);
+    GELOGD("Memory base changed, new memory base = %p", current_mem_base);
+    for (auto &task : tasks_) {
+      auto new_address = BuildTaskUtils::GetAddresses(task->GetOpdesc(), *running_param_);
+      GE_CHK_STATUS_RET(task->UpdateArgTable(*running_param_),
+                        "[%s] Failed to update arg table",
+                        task->GetOpdesc()->GetName().c_str());
+    }
+  }
   ret = UpdateArgs(inputs, outputs);
   if (ret != SUCCESS) {
     return ret;
@@ -225,9 +208,6 @@ DynamicSingleOp::DynamicSingleOp(uintptr_t resource_id, std::mutex *stream_mutex
     : resource_id_(resource_id), stream_mutex_(stream_mutex), stream_(stream) {
 }
 
-DynamicSingleOp::~DynamicSingleOp() {
-}
-
 Status DynamicSingleOp::ValidateParams(const vector<GeTensorDesc> &input_desc,
                                        const std::vector<DataBuffer> &inputs,
                                        std::vector<GeTensorDesc> &output_desc,
@@ -249,65 +229,24 @@ Status DynamicSingleOp::ValidateParams(const vector<GeTensorDesc> &input_desc,
   }
 
   if (input_desc.size() != num_inputs_) {
-    GELOGE(ACL_ERROR_GE_PARAM_INVALID, "Input number mismatches. expect %zu, but given %zu",
-           num_inputs_, input_desc.size());
+    GELOGE(ACL_ERROR_GE_PARAM_INVALID,
+           "Input number mismatches. expect %zu, but given %zu",
+           num_inputs_,
+           input_desc.size());
     return ACL_ERROR_GE_PARAM_INVALID;
   }
 
   if (output_desc.size() != num_outputs_) {
-    GELOGE(ACL_ERROR_GE_PARAM_INVALID, "Output number mismatches. expect %zu, but given %zu",
-           num_outputs_, output_desc.size());
+    GELOGE(ACL_ERROR_GE_PARAM_INVALID,
+           "Output number mismatches. expect %zu, but given %zu",
+           num_outputs_,
+           output_desc.size());
     return ACL_ERROR_GE_PARAM_INVALID;
   }
 
   return SUCCESS;
 }
 
-Status DynamicSingleOp::AllocateWorkspaces(const std::vector<int64_t> &workspace_sizes,
-                                           std::vector<void *> &workspaces) {
-  static const std::string kPurpose("malloc workspace memory for dynamic op.");
-  if (workspace_sizes.empty()) {
-    GELOGD("No need to allocate workspace.");
-    return SUCCESS;
-  }
-  int64_t total_size = 0;
-  std::vector<int64_t> ws_offsets;
-  for (auto ws_size : workspace_sizes) {
-    // alignment and padding should be done in OpParaCalculate
-    GE_CHK_STATUS_RET_NOLOG(CheckInt64AddOverflow(total_size, ws_size));
-    ws_offsets.emplace_back(total_size);
-    total_size += ws_size;
-  }
-
-  GELOGD("Total workspace size is %ld", total_size);
-  StreamResource *stream_resource = SingleOpManager::GetInstance().GetResource(resource_id_, stream_);
-  GE_CHECK_NOTNULL(stream_resource);
-  auto ws_base = stream_resource->MallocMemory(kPurpose, static_cast<size_t>(total_size));
-  if (ws_base == nullptr) {
-    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to allocate memory of size: %ld", total_size);
-    return ACL_ERROR_GE_MEMORY_ALLOCATION;
-  }
-  GELOGD("Done allocating workspace memory successfully.");
-
-  for (auto ws_offset : ws_offsets) {
-    workspaces.emplace_back(ws_base + ws_offset);
-  }
-
-  return SUCCESS;
-}
-
-Status DynamicSingleOp::ExecuteTbeTask(const vector<GeTensorDesc> &input_desc,
-                                       const vector<void *> &inputs,
-                                       vector<GeTensorDesc> &output_desc,
-                                       vector<void *> &outputs) {
-  GE_CHK_STATUS_RET_NOLOG(op_task_->UpdateRunInfo(input_desc, output_desc));
-
-  std::vector<void *> workspace_buffers;
-  GE_CHK_STATUS_RET_NOLOG(AllocateWorkspaces(op_task_->GetWorkspaceSizes(), workspace_buffers));
-
-  return op_task_->LaunchKernel(inputs, outputs, workspace_buffers, stream_);
-}
-
 Status DynamicSingleOp::ExecuteAsync(const vector<GeTensorDesc> &input_desc,
                                      const vector<DataBuffer> &input_buffers,
                                      vector<GeTensorDesc> &output_desc,
@@ -316,32 +255,8 @@ Status DynamicSingleOp::ExecuteAsync(const vector<GeTensorDesc> &input_desc,
   GE_CHK_STATUS_RET_NOLOG(ValidateParams(input_desc, input_buffers, output_desc, output_buffers));
   std::lock_guard<std::mutex> lk(*stream_mutex_);
 
-  std::vector<void *> inputs;
-  std::vector<void *> outputs;
-  for (auto &buffer : input_buffers) {
-    inputs.emplace_back(buffer.data);
-  }
-  for (auto &buffer : output_buffers) {
-    outputs.emplace_back(buffer.data);
-  }
-
-  if (op_task_->GetOpTaskType() == OP_TASK_TBE) {
-    auto ret = ExecuteTbeTask(input_desc, inputs, output_desc, outputs);
-    if (ret == SUCCESS) {
-      GE_CHK_STATUS_RET_NOLOG(ProfilingTaskInfo(op_task_.get()));
-    }
-    return ret;
-  } else if (op_task_->GetOpTaskType() == OP_TASK_AICPU || op_task_->GetOpTaskType() == OP_TASK_AICPUCC) {
-    auto aicpu_ret = op_task_->LaunchKernel(input_desc, input_buffers, output_desc, output_buffers, stream_);
-    if (aicpu_ret == SUCCESS) {
-      GE_CHK_STATUS_RET_NOLOG(ProfilingTaskInfo(op_task_.get()));
-    }
-    return aicpu_ret;
-  } else {
-    GELOGE(ACL_ERROR_GE_OP_TASK_TYPE_INVALID,
-           "Only TBE_Task, AI_CPU_Task and AI_CPUCC_Task are supported, but got %u",
-           op_task_->GetOpTaskType());
-    return ACL_ERROR_GE_OP_TASK_TYPE_INVALID;
-  }
+  GE_CHK_STATUS_RET_NOLOG(op_task_->LaunchKernel(input_desc, input_buffers, output_desc, output_buffers, stream_));
+  GE_CHK_STATUS_RET_NOLOG(ProfilingTaskInfo(op_task_.get()));
+  return SUCCESS;
 }
 }  // namespace ge

ge/single_op/single_op.h

@@ -30,9 +30,11 @@
 #include "cce/aicpu_engine_struct.h"
 
 namespace ge {
+class StreamResource;
+struct SingleOpModelParam;
 class SingleOp {
  public:
-  SingleOp(std::mutex *stream_mutex, rtStream_t stream);
+  SingleOp(StreamResource *stream_resource, std::mutex *stream_mutex, rtStream_t stream);
   ~SingleOp();
 
   Status ExecuteAsync(const std::vector<DataBuffer> &inputs, const std::vector<DataBuffer> &outputs);
@@ -44,6 +46,7 @@ class SingleOp {
   Status GetArgs(const std::vector<DataBuffer> &inputs, const std::vector<DataBuffer> &outputs);
 
   friend class SingleOpModel;
+  StreamResource *stream_resource_;
   std::mutex *stream_mutex_;
   rtStream_t stream_ = nullptr;
   std::vector<void *> input_addr_list_;
@@ -54,12 +57,13 @@ class SingleOp {
 
   std::vector<OpTask *> tasks_;
   std::vector<std::vector<uintptr_t *>> arg_table_;
+  std::unique_ptr<SingleOpModelParam> running_param_;
 };
 
 class DynamicSingleOp {
  public:
   DynamicSingleOp(uintptr_t resource_id, std::mutex *stream_mutex_, rtStream_t stream);
-  ~DynamicSingleOp();
+  ~DynamicSingleOp() = default;
   Status ExecuteAsync(const vector<GeTensorDesc> &input_desc,
                       const std::vector<DataBuffer> &inputs,
                       std::vector<GeTensorDesc> &output_desc,
@@ -72,14 +76,6 @@ class DynamicSingleOp {
                         std::vector<GeTensorDesc> &output_desc,
                         std::vector<DataBuffer> &outputs) const;
 
-  Status AllocateWorkspaces(const std::vector<int64_t> &workspace_sizes,
-                            std::vector<void *> &workspaces);
-
-  Status ExecuteTbeTask(const vector<GeTensorDesc> &input_desc,
-                        const vector<void *> &inputs,
-                        vector<GeTensorDesc> &output_desc,
-                        vector<void *> &outputs);
-
   std::unique_ptr<OpTask> op_task_;
   uintptr_t resource_id_ = 0;
   std::mutex *stream_mutex_;

ge/single_op/single_op_model.cc

@@ -92,7 +92,8 @@ Status SingleOpModel::InitModelMem(StreamResource &res) {
   if (model_params_.memory_size > model_params_.zero_copy_mem_size) {
     const string purpose("malloc feature map memory on model execute.");
     GELOGI("total memory: %lu, zero_copy_mem: %lu", model_params_.memory_size, model_params_.zero_copy_mem_size);
-    model_params_.mem_base = res.MallocMemory(purpose, model_params_.memory_size - model_params_.zero_copy_mem_size);
+    model_params_.mem_base =
+        res.MallocMemory(purpose, model_params_.memory_size - model_params_.zero_copy_mem_size, false);
     if (model_params_.mem_base == nullptr) {
       return ACL_ERROR_GE_MEMORY_ALLOCATION;
     }
@@ -226,9 +227,10 @@ Status SingleOpModel::SetInputsAndOutputs(SingleOp &single_op) {
   return SUCCESS;
 }
 
-Status SingleOpModel::BuildTaskList(SingleOp &single_op) {
+Status SingleOpModel::BuildTaskList(StreamResource *stream_resource, SingleOp &single_op) {
   auto ge_model = model_helper_.GetGeModel();
   GE_CHECK_NOTNULL(ge_model);
+  single_op.arg_table_.resize(single_op.input_sizes_.size() + single_op.output_sizes_.size());
   auto tasks = ge_model->GetModelTaskDefPtr()->task();
   for (int i = 0; i < tasks.size(); ++i) {
     const TaskDef &task_def = tasks[i];
@@ -247,9 +249,11 @@ Status SingleOpModel::BuildTaskList(SingleOp &single_op) {
           return ret;
         }
 
-        single_op.arg_table_.resize(single_op.input_sizes_.size() + single_op.output_sizes_.size());
         ParseArgTable(tbe_task, single_op);
         tbe_task->SetModelArgs(model_name_, model_id_);
+        if (tbe_task->tiling_buffer_ != nullptr) {
+          tbe_task->stream_resource_ = stream_resource;
+        }
         single_op.tasks_.emplace_back(tbe_task);
       } else if (kernel_type == ccKernelType::AI_CPU || kernel_type == ccKernelType::CUST_AI_CPU) {
         GELOGD("Building AICPU_CC task");
@@ -261,6 +265,7 @@ Status SingleOpModel::BuildTaskList(SingleOp &single_op) {
           return ret;
         }
         task->SetModelArgs(model_name_, model_id_);
+        ParseArgTable(task, single_op);
         single_op.tasks_.emplace_back(task);
       } else {
         GELOGE(ACL_ERROR_GE_OP_KERNEL_TYPE_INVALID,
@@ -278,6 +283,7 @@ Status SingleOpModel::BuildTaskList(SingleOp &single_op) {
         return ret;
       }
       aicpu_task->SetModelArgs(model_name_, model_id_);
+      ParseArgTable(aicpu_task, single_op);
       single_op.tasks_.emplace_back(aicpu_task);
     } else {
       // skip
@@ -287,21 +293,23 @@ Status SingleOpModel::BuildTaskList(SingleOp &single_op) {
   return SUCCESS;
 }
 
-void SingleOpModel::ParseArgTable(TbeOpTask *task, SingleOp &op) {
+void SingleOpModel::ParseArgTable(OpTask *task, SingleOp &op) {
   if (task == nullptr) {
     GELOGE(ACL_ERROR_GE_INTERNAL_ERROR, "tbe op task is nullptr");
     return;
   }
+
   // args: addr1, addr2, addr3 ...
-  auto *args = const_cast<uintptr_t *>(reinterpret_cast<const uintptr_t *>(task->GetArgs()));
-  size_t arg_size = task->GetArgSize();
-  for (size_t i = 0; i < arg_size / sizeof(void *); ++i) {
-    uintptr_t *ptr_to_addr = args + i;
+  uintptr_t *arg_base = nullptr;
+  size_t arg_num = 0;
+  task->GetIoAddr(arg_base, arg_num);
+  for (size_t i = 0; i < arg_num; ++i) {
+    uintptr_t *ptr_to_addr = arg_base + i;
     uintptr_t addr = *ptr_to_addr;
     auto iter = model_params_.addr_mapping_.find(addr);
     if (iter != model_params_.addr_mapping_.end()) {
       int arg_index = iter->second;
-      GELOGI("%s args[%zu] mapped to user designated args[%d]", task->GetStubName().c_str(), i, arg_index);
+      GELOGI("%s args[%zu] mapped to user designated args[%d]", task->GetOpdesc()->GetName().c_str(), i, arg_index);
       op.arg_table_[iter->second].emplace_back(ptr_to_addr);
     }
   }
@@ -386,8 +394,10 @@ Status SingleOpModel::BuildCpuKernelTask(const domi::KernelDef &kernel_def, OpTa
 Status SingleOpModel::BuildOp(StreamResource &resource, SingleOp &single_op) {
   GE_CHK_STATUS_RET_NOLOG(ParseInputsAndOutputs());
   GE_CHK_STATUS_RET_NOLOG(InitModelMem(resource));
+  single_op.running_param_.reset(new (std::nothrow)SingleOpModelParam(model_params_));
+  GE_CHECK_NOTNULL(single_op.running_param_);
   GE_CHK_STATUS_RET_NOLOG(SetInputsAndOutputs(single_op));
-  return BuildTaskList(single_op);
+  return BuildTaskList(&resource, single_op);
 }
 
 Status SingleOpModel::BuildModelTaskKernel(const TaskDef &task_def, DynamicSingleOp &single_op) {

ge/single_op/single_op_model.h

@@ -65,7 +65,7 @@ class SingleOpModel {
   Status ParseInputNode(const OpDescPtr &op_desc);
   void ParseOutputNode(const OpDescPtr &op_desc);
 
-  Status BuildTaskList(SingleOp &single_op);
+  Status BuildTaskList(StreamResource *stream_resource, SingleOp &single_op);
   Status BuildTaskListForDynamicOp(DynamicSingleOp &dynamic_single_op);
   Status BuildKernelTask(const domi::KernelDef &kernel_def, TbeOpTask **task);
   Status BuildKernelExTask(const domi::KernelExDef &kernel_def, AiCpuTask **task,
@@ -74,7 +74,7 @@ class SingleOpModel {
   Status BuildModelTaskKernel(const domi::TaskDef &task_def, DynamicSingleOp &single_op);
 
   static void ParseOpModelParams(ModelHelper &model_helper, SingleOpModelParam &param);
-  void ParseArgTable(TbeOpTask *task, SingleOp &op);
+  void ParseArgTable(OpTask *task, SingleOp &op);
 
   std::string model_name_;
   uint32_t model_id_ = 0;

ge/single_op/stream_resource.cc

@@ -69,11 +69,25 @@ uint8_t *StreamResource::DoMallocMemory(const std::string &purpose,
                                         size_t size,
                                         size_t &max_allocated,
                                         std::vector<uint8_t *> &allocated) {
+  if (size == 0) {
+    GELOGD("Mem size == 0");
+    return nullptr;
+  }
+
   if (size <= max_allocated && !allocated.empty()) {
     GELOGD("reuse last memory");
     return allocated.back();
   }
 
+  if (!allocated.empty()) {
+    uint8_t *current_buffer = allocated.back();
+    allocated.pop_back();
+    if (rtStreamSynchronize(stream_) != RT_ERROR_NONE) {
+      GELOGW("Failed to invoke rtStreamSynchronize");
+    }
+    (void) rtFree(current_buffer);
+  }
+
   uint8_t *buffer = nullptr;
   auto ret = rtMalloc(reinterpret_cast<void **>(&buffer), size, RT_MEMORY_HBM);
   if (ret != RT_ERROR_NONE) {
@@ -96,10 +110,14 @@ uint8_t *StreamResource::DoMallocMemory(const std::string &purpose,
   return buffer;
 }
 
-uint8_t *StreamResource::MallocMemory(const std::string &purpose, size_t size) {
+uint8_t *StreamResource::MallocMemory(const std::string &purpose, size_t size, bool holding_lock) {
   GELOGD("To Malloc memory, size = %zu", size);
-  uint8_t *buffer = DoMallocMemory(purpose, size, max_memory_size_, memory_list_);
-  return buffer;
+  if (holding_lock) {
+    return DoMallocMemory(purpose, size, max_memory_size_, memory_list_);
+  } else {
+    std::lock_guard<std::mutex> lk(stream_mu_);
+    return DoMallocMemory(purpose, size, max_memory_size_, memory_list_);
+  }
 }
 
 uint8_t *StreamResource::MallocWeight(const std::string &purpose, size_t size) {
@@ -158,7 +176,7 @@ Status StreamResource::BuildOperator(const string &model_name, const ModelData &
     return ret;
   }
 
-  auto new_op = std::unique_ptr<SingleOp>(new(std::nothrow) SingleOp(&stream_mu_, stream_));
+  auto new_op = std::unique_ptr<SingleOp>(new(std::nothrow) SingleOp(this, &stream_mu_, stream_));
   if (new_op == nullptr) {
     GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "new SingleOp failed");
     return ACL_ERROR_GE_MEMORY_ALLOCATION;
@@ -171,4 +189,12 @@ Status StreamResource::BuildOperator(const string &model_name, const ModelData &
   op_map_[model_data.model_data] = std::move(new_op);
   return SUCCESS;
 }
+
+const uint8_t *StreamResource::GetMemoryBase() const {
+  if (memory_list_.empty()) {
+    return nullptr;
+  }
+
+  return memory_list_.back();
+}
 }  // namespace ge

ge/single_op/stream_resource.h

@@ -45,8 +45,9 @@ class StreamResource {
   Status BuildOperator(const std::string &model_name, const ModelData &model_data, SingleOp **single_op);
   Status BuildDynamicOperator(const std::string &model_name, const ModelData &model_data, DynamicSingleOp **single_op);
 
-  uint8_t *MallocMemory(const std::string &purpose, size_t size);
+  uint8_t *MallocMemory(const std::string &purpose, size_t size, bool holding_lock = true);
   uint8_t *MallocWeight(const std::string &purpose, size_t size);
+  const uint8_t *GetMemoryBase() const;
 
  private:
   uint8_t *DoMallocMemory(const std::string &purpose,

ge/single_op/task/aicpu_kernel_task_builder.cc

@@ -17,17 +17,22 @@
 #include "single_op/task/aicpu_kernel_task_builder.h"
 #include "framework/common/taskdown_common.h"
 #include "graph/load/new_model_manager/model_manager.h"
+#include "build_task_utils.h"
 
 namespace ge {
 AiCpuCCTaskBuilder::AiCpuCCTaskBuilder(const OpDescPtr &op_desc, const domi::KernelDef &kernel_def)
     : op_desc_(op_desc), kernel_def_(kernel_def) {}
 
-Status AiCpuCCTaskBuilder::SetKernelArgs(AiCpuCCTask &task) {
+Status AiCpuCCTaskBuilder::SetKernelArgs(AiCpuCCTask &task, const SingleOpModelParam &param) {
   size_t aicpu_arg_size = kernel_def_.args_size();
-  if (aicpu_arg_size <= 0) {
+  if (aicpu_arg_size <= sizeof(aicpu::AicpuParamHead)) {
     GELOGE(ACL_ERROR_GE_PARAM_INVALID, "aicpu_arg_size is invalid, value = %zu", aicpu_arg_size);
     return ACL_ERROR_GE_PARAM_INVALID;
   }
+
+  task.io_addr_num_ = op_desc_->GetInputsSize() + op_desc_->GetOutputsSize();
+  GE_CHECK_GE(aicpu_arg_size - sizeof(aicpu::AicpuParamHead), task.io_addr_num_ * sizeof(void *));
+
   std::unique_ptr<uint8_t[]> aicpu_args;
   aicpu_args.reset(new(std::nothrow) uint8_t[aicpu_arg_size]());
   if (aicpu_args == nullptr) {
@@ -41,13 +46,19 @@ Status AiCpuCCTaskBuilder::SetKernelArgs(AiCpuCCTask &task) {
     return ACL_ERROR_GE_INTERNAL_ERROR;
   }
 
-  task.SetIoAddr(aicpu_args.get() + sizeof(aicpu::AicpuParamHead));
+  task.SetIoAddr(reinterpret_cast<uintptr_t *>(aicpu_args.get() + sizeof(aicpu::AicpuParamHead)));
   task.SetKernelArgs(std::move(aicpu_args), aicpu_arg_size);
+
+  auto addresses = BuildTaskUtils::GetKernelArgs(op_desc_, param);
+  GE_CHECK_GE(addresses.size(), task.io_addr_num_);
+  for (size_t i = 0; i < task.io_addr_num_; ++i) {
+    task.io_addr_[i] = reinterpret_cast<uintptr_t>(addresses[i]);
+  }
   return SUCCESS;
 }
 
-Status AiCpuCCTaskBuilder::BuildTask(AiCpuCCTask &task, uint64_t kernel_id) {
-  auto ret = SetKernelArgs(task);
+Status AiCpuCCTaskBuilder::BuildTask(AiCpuCCTask &task, uint64_t kernel_id, const SingleOpModelParam &param) {
+  auto ret = SetKernelArgs(task, param);
   if (ret != SUCCESS) {
     return ret;
   }

ge/single_op/task/aicpu_kernel_task_builder.h

@@ -30,10 +30,10 @@ class AiCpuCCTaskBuilder {
   explicit AiCpuCCTaskBuilder(const OpDescPtr &op_desc, const domi::KernelDef &kernel_def);
   ~AiCpuCCTaskBuilder() = default;
 
-  Status BuildTask(AiCpuCCTask &task, uint64_t kernel_id);
+  Status BuildTask(AiCpuCCTask &task, uint64_t kernel_id, const SingleOpModelParam &param);
 
  private:
-  Status SetKernelArgs(AiCpuCCTask &task);
+  Status SetKernelArgs(AiCpuCCTask &task, const SingleOpModelParam &param);
   const OpDescPtr op_desc_;
   const domi::KernelDef &kernel_def_;
 };

ge/single_op/task/aicpu_task_builder.cc

@@ -26,26 +26,6 @@ namespace ge {
   AiCpuTaskBuilder::AiCpuTaskBuilder(const OpDescPtr &op_desc, const domi::KernelExDef &kernel_def)
       : op_desc_(op_desc), kernel_def_(kernel_def) {}
 
-  Status AiCpuTaskBuilder::SetInputOutputAddr(void **io_addr, const std::vector<void *> &addresses) {
-    size_t arg_size = kernel_def_.args_size();
-    auto rt_ret = rtMalloc(io_addr, arg_size, RT_MEMORY_HBM);
-    if (rt_ret != RT_ERROR_NONE) {
-      GELOGE(rt_ret, "rtMalloc failed, size = %zu, ret = %d", arg_size, rt_ret);
-      return rt_ret;
-    }
-
-    const void *src_addr = reinterpret_cast<const void *>(addresses.data());
-    uint64_t src_len = sizeof(void *) * addresses.size();
-    rt_ret = rtMemcpy(*io_addr, arg_size, src_addr, src_len, RT_MEMCPY_HOST_TO_DEVICE);
-    if (rt_ret != RT_ERROR_NONE) {
-      (void)rtFree(*io_addr);
-      GELOGE(rt_ret, "rtMemcpy addresses failed, ret = %d", rt_ret);
-      return rt_ret;
-    }
-
-    return SUCCESS;
-  }
-
   Status AiCpuTaskBuilder::SetFmkOpKernel(void *io_addr, void *ws_addr, STR_FWK_OP_KERNEL &fwk_op_kernel) {
     auto sec_ret = memcpy_s(&fwk_op_kernel, sizeof(STR_FWK_OP_KERNEL),
                             kernel_def_.args().data(), kernel_def_.args().size());
@@ -80,39 +60,27 @@ namespace ge {
     return SUCCESS;
   }
 
-  Status AiCpuTaskBuilder::InitWorkspaceAndIO(void **io_addr, void **kernel_workspace,
-                                              const SingleOpModelParam &param, bool dynamic_flag) {
+  Status AiCpuTaskBuilder::InitWorkspaceAndIO(AiCpuTask &task, const SingleOpModelParam &param, bool dynamic_flag) {
     if (kernel_def_.args_size() > sizeof(STR_FWK_OP_KERNEL)) {
       GELOGE(ACL_ERROR_GE_PARAM_INVALID, "sizeof STR_FWK_OP_KERNEL is: %lu, but args_size is: %d",
              sizeof(STR_FWK_OP_KERNEL), kernel_def_.args_size());
       return ACL_ERROR_GE_PARAM_INVALID;
     }
-    auto addresses = BuildTaskUtils::GetAddresses(op_desc_, param);
-    auto ws_addr_vec = addresses.at(BuildTaskUtils::kAddressIndexWorkspace);
-
-    if (dynamic_flag) {
-      GE_CHK_RT_RET(rtMalloc(kernel_workspace, kernel_def_.task_info_size(), RT_MEMORY_HBM));
-    } else {
-      if (ws_addr_vec.empty()) {
-        GELOGE(ACL_ERROR_GE_PARAM_INVALID, "workspace Data Address is empty.");
-        return ACL_ERROR_GE_PARAM_INVALID;
-      }
-      *kernel_workspace = ws_addr_vec[0];
-    }
-    GE_CHK_RT_RET(rtMemcpy(*kernel_workspace, kernel_def_.task_info_size(),
+    GE_CHK_RT_RET(rtMalloc(&task.workspace_addr_, kernel_def_.task_info_size(), RT_MEMORY_HBM));
+    GE_CHK_RT_RET(rtMemcpy(task.workspace_addr_, kernel_def_.task_info_size(),
                            kernel_def_.task_info().data(), kernel_def_.task_info_size(),
                            RT_MEMCPY_HOST_TO_DEVICE));
 
-    auto ret = SetInputOutputAddr(io_addr, BuildTaskUtils::JoinAddresses(addresses));
-    if (ret != SUCCESS) {
-      return ret;
-    }
+    auto addresses = BuildTaskUtils::GetAddresses(op_desc_, param, false);
+    task.io_addr_host_ = BuildTaskUtils::JoinAddresses(addresses);
+    task.io_addr_size_ = task.io_addr_host_.size() * sizeof(void *);
+    GE_CHK_RT_RET(rtMalloc(&task.io_addr_, task.io_addr_size_, RT_MEMORY_HBM));
     return SUCCESS;
   }
 
   Status AiCpuTaskBuilder::BuildTask(ge::AiCpuTask &task, const SingleOpModelParam &param,
                                      bool dynamic_flag, uint64_t kernel_id) {
-    GE_CHK_STATUS_RET_NOLOG(InitWorkspaceAndIO(&task.io_addr_, &task.workspace_addr_, param, dynamic_flag));
+    GE_CHK_STATUS_RET_NOLOG(InitWorkspaceAndIO(task, param, dynamic_flag));
 
     STR_FWK_OP_KERNEL fwk_op_kernel = {0};
     auto ret = SetFmkOpKernel(task.io_addr_, task.workspace_addr_, fwk_op_kernel);

ge/single_op/task/aicpu_task_builder.h

@@ -33,10 +33,8 @@ namespace ge {
 
   private:
     static Status SetKernelArgs(void **args, STR_FWK_OP_KERNEL &kernel);
-    Status SetInputOutputAddr(void **io_addr, const std::vector<void *> &addresses);
     Status SetFmkOpKernel(void *io_addr, void *ws_addr, STR_FWK_OP_KERNEL &kernel);
-    Status InitWorkspaceAndIO(void **io_addr, void **kernel_workspace,
-                              const SingleOpModelParam &param, bool dynamic_flag);
+    Status InitWorkspaceAndIO(AiCpuTask &task, const SingleOpModelParam &param, bool dynamic_flag);
 
     const OpDescPtr op_desc_;
     const domi::KernelExDef &kernel_def_;

ge/single_op/task/build_task_utils.cc

@@ -32,7 +32,8 @@ const uint64_t kVarSize = 0;
 }
 
 std::vector<std::vector<void *>> BuildTaskUtils::GetAddresses(const OpDescPtr &op_desc,
-                                                              const SingleOpModelParam &param) {
+                                                              const SingleOpModelParam &param,
+                                                              bool keep_workspace) {
   std::vector<std::vector<void *>> ret;
   RuntimeParam runtime_para;
   runtime_para.mem_size = param.memory_size;
@@ -49,7 +50,9 @@ std::vector<std::vector<void *>> BuildTaskUtils::GetAddresses(const OpDescPtr &o
 
   ret.emplace_back(ModelUtils::GetInputDataAddrs(runtime_para, op_desc));
   ret.emplace_back(ModelUtils::GetOutputDataAddrs(runtime_para, op_desc));
-  ret.emplace_back(ModelUtils::GetWorkspaceDataAddrs(runtime_para, op_desc));
+  if (keep_workspace) {
+    ret.emplace_back(ModelUtils::GetWorkspaceDataAddrs(runtime_para, op_desc));
+  }
   return ret;
 }
 

ge/single_op/task/build_task_utils.h

@@ -27,15 +27,17 @@
 namespace ge {
 class BuildTaskUtils {
  public:
+  static constexpr int kAddressIndexOutput = 1;
   static constexpr int kAddressIndexWorkspace = 2;
 
-  static std::vector<std::vector<void *>> GetAddresses(const OpDescPtr &op_desc, const SingleOpModelParam &param);
+  static std::vector<std::vector<void *>> GetAddresses(const OpDescPtr &op_desc,
+                                                       const SingleOpModelParam &param,
+                                                       bool keep_workspace = true);
   static std::vector<void *> JoinAddresses(const std::vector<std::vector<void *>> &addresses);
   static std::vector<void *> GetKernelArgs(const OpDescPtr &op_desc, const SingleOpModelParam &param);
   static std::string GetTaskInfo(const OpDescPtr &op_desc);
   template<typename T>
-  static std::string VectorToString(const std::vector<T> &values)
-  {
+  static std::string VectorToString(const std::vector<T> &values) {
     std::stringstream ss;
     ss << '[';
     auto size = values.size();

ge/single_op/task/op_task.cc

@@ -24,9 +24,11 @@
 #include "common/dump/dump_manager.h"
 #include "common/dump/dump_op.h"
 #include "common/formats/formats.h"
+#include "common/math/math_util.h"
 #include "framework/common/debug/log.h"
 #include "register/op_tiling.h"
 #include "runtime/rt.h"
+#include "build_task_utils.h"
 
 namespace ge {
 namespace {
@@ -48,18 +50,22 @@ Status OpTask::OpenDump(rtStream_t stream) {
     std::vector<uint64_t> output_adds;
     auto input_size = op_desc_->GetInputsSize();
     auto output_size = op_desc_->GetOutputsSize();
-    auto all_size = io_addrs_for_dump_.size();
-    if (input_size + output_size != all_size) {
-      GELOGE(FAILED, "io_addrs_for_dump_ size %zu is not equal input and output size %zu", all_size,
+    uintptr_t *arg_base = nullptr;
+    size_t arg_num = 0;
+    GetIoAddr(arg_base, arg_num);
+    if (arg_num < input_size + output_size) {
+      GELOGE(FAILED, "io_addrs_for_dump_ size %zu is not equal input and output size %zu",
+             arg_num,
              input_size + output_size);
       return FAILED;
     }
+
     for (size_t i = 0; i < input_size; i++) {
-      uint64_t input_addr = io_addrs_for_dump_[i];
+      uint64_t input_addr = arg_base[i];
       input_addrs.emplace_back(input_addr);
     }
     for (size_t j = 0; j < output_size; j++) {
-      uint64_t output_addr = io_addrs_for_dump_[input_size + j];
+      uint64_t output_addr = arg_base[input_size + j];
       output_adds.emplace_back(output_addr);
     }
     dump_op_.SetDumpInfo(DumpManager::GetInstance().GetDumpProperties(), op_desc_, input_addrs, output_adds, stream);
@@ -89,10 +95,6 @@ void TbeOpTask::SetKernelArgs(std::unique_ptr<uint8_t[]> &&args, size_t arg_size
 
 void TbeOpTask::SetSmDesc(void *sm_desc) { sm_desc_ = sm_desc; }
 
-const vector<int64_t> &OpTask::GetWorkspaceSizes() const { return workspace_sizes_; }
-
-void OpTask::SetWorkspaceSizes(const vector<int64_t> &workspace_sizes) { workspace_sizes_ = workspace_sizes; }
-
 void OpTask::SetModelArgs(std::string model_name, uint32_t model_id) {
   model_name_ = model_name;
   model_id_ = model_id;
@@ -107,6 +109,36 @@ Status OpTask::GetProfilingArgs(std::string &model_name, std::string &op_name, u
   op_name = op_desc_->GetName();
   return SUCCESS;
 }
+Status OpTask::UpdateRunInfo(const vector<GeTensorDesc> &input_desc, const vector<GeTensorDesc> &output_desc) {
+  return UNSUPPORTED;
+}
+Status OpTask::UpdateArgTable(const SingleOpModelParam &param) {
+  auto addresses = BuildTaskUtils::GetAddresses(op_desc_, param);
+  auto all_addresses = BuildTaskUtils::JoinAddresses(addresses);
+  uintptr_t *arg_base = nullptr;
+  size_t arg_num = 0;
+  GetIoAddr(arg_base, arg_num);
+  if (arg_num !=  all_addresses.size()) {
+    GELOGE(INTERNAL_ERROR, "[%s] arg number mismatches, expect = %zu, but got = %zu",
+           op_desc_->GetName().c_str(),
+           arg_num,
+           all_addresses.size());
+    return INTERNAL_ERROR;
+  }
+
+  for (void *addr : all_addresses) {
+    *arg_base++ = reinterpret_cast<uintptr_t >(addr);
+  }
+  return SUCCESS;
+}
+
+Status OpTask::LaunchKernel(const vector<GeTensorDesc> &input_desc,
+                            const vector<DataBuffer> &input_buffers,
+                            vector<GeTensorDesc> &output_desc,
+                            vector<DataBuffer> &output_buffers,
+                            rtStream_t stream) {
+  return UNSUPPORTED;
+}
 
 TbeOpTask::~TbeOpTask() {
   if (sm_desc_ != nullptr) {
@@ -141,12 +173,6 @@ Status TbeOpTask::LaunchKernel(rtStream_t stream) {
     return RT_FAILED;
   }
   GELOGI("[TASK_INFO] %s", this->stub_name_.c_str());
-
-  size_t input_size = op_desc_->GetInputsSize();
-  size_t output_size = op_desc_->GetOutputsSize();
-  uint64_t *io_addr = reinterpret_cast<uint64_t *>(args_.get());
-  std::vector<uint64_t> io_addrs(io_addr, io_addr + input_size + output_size);
-  SetIoAddrsForDump(io_addrs);
   auto status = OpenDump(stream);
   if (status != SUCCESS) {
     GELOGE(status, "Open dump failed in the tbe single op %s", this->stub_name_.c_str());
@@ -167,11 +193,12 @@ Status TbeOpTask::UpdateRunInfo(const vector<GeTensorDesc> &input_desc, const ve
     GELOGE(FAILED, "Failed to invoke OpParaCalculate. ret = %u", ret);
     return FAILED;
   }
-  SetWorkspaceSizes(run_info.workspaces);
   block_dim_ = run_info.block_dim;
   tiling_data_ = run_info.tiling_data.str();
   GELOGD("Done invoking OpParaCalculate successfully. block_dim = %u, tiling size = %zu", block_dim_,
          tiling_data_.size());
+
+  GE_CHK_STATUS_RET(AllocateWorkspaces(run_info.workspaces), "Failed to allocate workspaces");
   return SUCCESS;
 }
 
@@ -227,13 +254,54 @@ void TbeOpTask::EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, s
   max_tiling_size_ = max_tiling_size;
 }
 
-Status TbeOpTask::LaunchKernel(const vector<void *> &inputs, const vector<void *> &outputs,
-                               const vector<void *> &workspaces, rtStream_t stream) {
+Status TbeOpTask::AllocateWorkspaces(const vector<int64_t> &workspace_sizes) {
+  static const std::string kPurpose("malloc workspace memory for dynamic op.");
+  if (workspace_sizes.empty()) {
+    GELOGD("No need to allocate workspace.");
+    return SUCCESS;
+  }
+  int64_t total_size = 0;
+  std::vector<int64_t> ws_offsets;
+  for (auto ws_size : workspace_sizes) {
+    // alignment and padding should be done in OpParaCalculate
+    GE_CHK_STATUS_RET_NOLOG(CheckInt64AddOverflow(total_size, ws_size));
+    ws_offsets.emplace_back(total_size);
+    total_size += ws_size;
+  }
+
+  GELOGD("Total workspace size is %ld", total_size);
+  GE_CHECK_NOTNULL(stream_resource_);
+  auto ws_base = stream_resource_->MallocMemory(kPurpose, static_cast<size_t>(total_size));
+  if (ws_base == nullptr) {
+    GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to allocate memory of size: %ld", total_size);
+    return ACL_ERROR_GE_MEMORY_ALLOCATION;
+  }
+  GELOGD("Done allocating workspace memory successfully.");
+
+  for (auto ws_offset : ws_offsets) {
+    workspaces_.emplace_back(ws_base + ws_offset);
+  }
+
+  return SUCCESS;
+}
+
+Status TbeOpTask::LaunchKernel(const vector<GeTensorDesc> &input_desc,
+                               const vector<DataBuffer> &input_buffers,
+                               vector<GeTensorDesc> &output_desc,
+                               vector<DataBuffer> &output_buffers,
+                               rtStream_t stream) {
+  GE_CHK_STATUS_RET_NOLOG(UpdateRunInfo(input_desc, output_desc));
   GELOGD("[%s] Start to launch kernel", node_->GetName().c_str());
   std::vector<void *> args;
-  args.insert(args.end(), inputs.begin(), inputs.end());
-  args.insert(args.end(), outputs.begin(), outputs.end());
-  args.insert(args.end(), workspaces.begin(), workspaces.end());
+  for (auto &buffer : input_buffers) {
+    args.emplace_back(buffer.data);
+  }
+  for (auto &buffer : output_buffers) {
+    args.emplace_back(buffer.data);
+  }
+  for (auto &buffer : workspaces_) {
+    args.emplace_back(buffer);
+  }
 
   if (tiling_buffer_ != nullptr) {
     GELOGD("[%s] Start to copy tiling info. size = %zu", node_->GetName().c_str(), tiling_data_.size());
@@ -254,6 +322,14 @@ Status TbeOpTask::LaunchKernel(const vector<void *> &inputs, const vector<void *
   return SUCCESS;
 }
 
+void TbeOpTask::GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) {
+  arg_base = reinterpret_cast<uintptr_t *>(args_.get());
+  arg_count = arg_size_ / sizeof(void *);
+  if (tiling_buffer_ != nullptr) {
+    --arg_count;
+  }
+}
+
 AiCpuBaseTask::~AiCpuBaseTask() {
   if (ext_info_addr_dev_ != nullptr) {
     (void)rtFree(ext_info_addr_dev_);
@@ -399,12 +475,14 @@ AiCpuTask::~AiCpuTask() {
   }
 }
 
-const void *AiCpuTask::GetIOAddr() const { return io_addr_; }
-
 Status AiCpuTask::LaunchKernel(rtStream_t stream) {
   GELOGD("Start to launch kernel. task = %s", this->op_type_.c_str());
-  auto ret = rtMemcpyAsync(workspace_addr_, task_info_.size(), task_info_.data(), task_info_.size(),
-                           RT_MEMCPY_HOST_TO_DEVICE_EX, stream);
+  auto ret = rtMemcpyAsync(io_addr_,
+                           io_addr_size_,
+                           io_addr_host_.data(),
+                           io_addr_host_.size() * sizeof(void *),
+                           RT_MEMCPY_HOST_TO_DEVICE_EX,
+                           stream);
   if (ret != RT_ERROR_NONE) {
     GELOGE(RT_FAILED, "rtMemcpyAsync workspace data failed. ret = %d, task = %s", ret, this->op_type_.c_str());
     return RT_FAILED;
@@ -680,6 +758,17 @@ Status AiCpuTask::LaunchKernel(const std::vector<GeTensorDesc> &input_desc,
   return SUCCESS;
 }
 
+Status AiCpuTask::UpdateArgTable(const SingleOpModelParam &param) {
+  auto addresses = BuildTaskUtils::GetAddresses(op_desc_, param, false);
+  io_addr_host_ = BuildTaskUtils::JoinAddresses(addresses);
+  return SUCCESS;
+}
+
+void AiCpuTask::GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) {
+  arg_base = reinterpret_cast<uintptr_t *>(io_addr_host_.data());
+  arg_count = io_addr_host_.size();
+}
+
 void AiCpuCCTask::SetKernelArgs(std::unique_ptr<uint8_t[]> args, size_t arg_size) {
   args_ = std::move(args);
   arg_size_ = arg_size;
@@ -691,9 +780,7 @@ void AiCpuCCTask::SetSoName(const std::string &so_name) { so_name_ = so_name; }
 
 void AiCpuCCTask::SetkernelName(const std::string &kernel_Name) { kernel_name_ = kernel_Name; }
 
-void AiCpuCCTask::SetIoAddr(void *io_addr) { io_addr_ = io_addr; }
-
-const void *AiCpuCCTask::GetIOAddr() const { return io_addr_; }
+void AiCpuCCTask::SetIoAddr(uintptr_t *io_addr) { io_addr_ = io_addr; }
 
 const void *AiCpuCCTask::GetArgs() const { return args_.get(); }
 
@@ -716,12 +803,6 @@ Status AiCpuCCTask::LaunchKernel(rtStream_t stream) {
     return ret;
   }
   GELOGD("Invoke rtCpuKernelLaunch succeeded");
-
-  size_t input_size = op_desc_->GetInputsSize();
-  size_t output_size = op_desc_->GetOutputsSize();
-  uint64_t *io_addr = reinterpret_cast<uint64_t *>(io_addr_);
-  std::vector<uint64_t> io_addrs (io_addr, io_addr + input_size + output_size);
-  SetIoAddrsForDump(io_addrs);
   auto status = OpenDump(stream);
   if (status != SUCCESS) {
     GELOGE(status, "Open dump failed in the aicpucc single op %s", this->kernel_name_.c_str());
@@ -761,4 +842,9 @@ Status AiCpuCCTask::LaunchKernel(const std::vector<GeTensorDesc> &input_desc,
 
   return SUCCESS;
 }
+
+void AiCpuCCTask::GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) {
+  arg_base = io_addr_;
+  arg_count = io_addr_num_;
+}
 }  // namespace ge

ge/single_op/task/op_task.h

@@ -32,49 +32,27 @@
 #include "init/gelib.h"
 
 namespace ge {
-enum OpTaskType {
-  OP_TASK_TBE = 0,
-  OP_TASK_AICPU,
-  OP_TASK_AICPUCC,
-  OP_TASK_INVALID,
-};
-
+class StreamResource;
+struct SingleOpModelParam;
 class OpTask {
  public:
   OpTask() = default;
   virtual ~OpTask() = default;
   virtual Status LaunchKernel(rtStream_t stream) = 0;
   virtual Status UpdateRunInfo(const vector<GeTensorDesc> &input_desc,
-                               const vector<GeTensorDesc> &output_desc) {
-    return UNSUPPORTED;
-  }
-  virtual Status LaunchKernel(const std::vector<void *> &inputs,
-                              const std::vector<void *> &outputs,
-                              const std::vector<void *> &workspaces,
-                              rtStream_t stream) {
-    return UNSUPPORTED;
-  }
-  virtual OpTaskType GetOpTaskType() = 0;
-  virtual const void *GetIOAddr() const = 0;
-  const vector<int64_t> &GetWorkspaceSizes() const;
-  void SetWorkspaceSizes(const vector<int64_t> &workspace_sizes);
+                               const vector<GeTensorDesc> &output_desc);
+  virtual Status UpdateArgTable(const SingleOpModelParam &param);
   void SetModelArgs(std::string model_name, uint32_t model_id);
   Status GetProfilingArgs(std::string &model_name, std::string &op_name, uint32_t &model_id, uint32_t &block_dim);
   const OpDescPtr &GetOpdesc() const {return op_desc_;}
   Status OpenDump(rtStream_t stream);
-  void SetIoAddrsForDump(const vector<uint64_t> &io_addrs_for_dump) {
-    io_addrs_for_dump_ = io_addrs_for_dump;
-  }
+  virtual void GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) = 0;
   virtual Status LaunchKernel(const std::vector<GeTensorDesc> &input_desc,
                               const std::vector<DataBuffer> &input_buffers,
                               std::vector<GeTensorDesc> &output_desc,
                               std::vector<DataBuffer> &output_buffers,
-                              rtStream_t stream) {
-    return UNSUPPORTED;
-  }
+                              rtStream_t stream);
 
- private:
-  std::vector<int64_t> workspace_sizes_;
  protected:
   DumpProperties dump_properties_;
   DumpOp dump_op_;
@@ -82,19 +60,18 @@ class OpTask {
   std::string model_name_;
   uint32_t model_id_ = 0;
   uint32_t block_dim_ = 1;
-  std::vector<uint64_t> io_addrs_for_dump_;
 };
 
 class TbeOpTask : public OpTask {
  public:
   ~TbeOpTask() override;
   Status LaunchKernel(rtStream_t stream) override;
-  OpTaskType GetOpTaskType() override {
-    return OP_TASK_TBE;
-  }
-  const void *GetIOAddr() const override {
-    return nullptr;
-  }
+  Status LaunchKernel(const std::vector<GeTensorDesc> &input_desc,
+                      const std::vector<DataBuffer> &input_buffers,
+                      std::vector<GeTensorDesc> &output_desc,
+                      std::vector<DataBuffer> &output_buffers,
+                      rtStream_t stream) override;
+  void GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) override;
   void SetSmDesc(void *sm_desc);
   void SetStubFunc(const std::string &name, const void *stub_func);
   void SetKernelArgs(std::unique_ptr<uint8_t[]> &&args, size_t arg_size, uint32_t block_dim, const OpDescPtr &op_desc);
@@ -102,20 +79,17 @@ class TbeOpTask : public OpTask {
   Status UpdateRunInfo(const vector<GeTensorDesc> &input_desc,
                        const vector<GeTensorDesc> &output_desc) override;
 
-  Status LaunchKernel(const vector<void *> &inputs,
-                      const vector<void *> &outputs,
-                      const vector<void *> &workspaces,
-                      rtStream_t stream) override;
-
   const void *GetArgs() const;
   size_t GetArgSize() const;
   const std::string &GetStubName() const;
   void EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, size_t max_tiling_size);
 
  private:
+  friend class SingleOpModel;
   static Status UpdateTensorDesc(const GeTensorDesc &src_tensor, GeTensorDesc &dst_tensor);
   Status UpdateNodeByShape(const vector<GeTensorDesc> &input_desc,
                            const vector<GeTensorDesc> &output_desc);
+  Status AllocateWorkspaces(const std::vector<int64_t> &workspace_sizes);
 
   const void *stub_func_ = nullptr;
   std::unique_ptr<uint8_t[]> args_;
@@ -123,9 +97,11 @@ class TbeOpTask : public OpTask {
   void *sm_desc_ = nullptr;
   std::string stub_name_;
 
+  StreamResource *stream_resource_ = nullptr;
   void *tiling_buffer_ = nullptr;
   uint32_t max_tiling_size_ = 0;
   std::string tiling_data_;
+  std::vector<void *> workspaces_;
   NodePtr node_;
 };
 
@@ -133,7 +109,7 @@ class AiCpuBaseTask : public OpTask {
  public:
   AiCpuBaseTask() = default;
   ~AiCpuBaseTask() override;
-  const UnknowShapeOpType GetUnknownType() const { return unknown_type_; }
+  UnknowShapeOpType GetUnknownType() const { return unknown_type_; }
 
  protected:
   Status SetExtInfoAndType(const std::string &kernel_ext_info, uint64_t kernel_id);
@@ -158,10 +134,8 @@ class AiCpuTask : public AiCpuBaseTask {
   ~AiCpuTask() override;
 
   Status LaunchKernel(rtStream_t stream) override;
-  OpTaskType GetOpTaskType() override {
-    return OP_TASK_AICPU;
-  }
-  const void *GetIOAddr() const override;
+  Status UpdateArgTable(const SingleOpModelParam &param) override;
+  void GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) override;
 
   Status LaunchKernel(const std::vector<GeTensorDesc> &input_desc,
                       const std::vector<DataBuffer> &input_buffers,
@@ -188,27 +162,31 @@ class AiCpuTask : public AiCpuBaseTask {
   friend class AiCpuTaskBuilder;
   void *workspace_addr_ = nullptr;
   std::string task_info_;
- // device addr
+  // device addr
   void *args_ = nullptr;
   size_t arg_size_ = 0;
   std::string op_type_;
   // device addr
   void *io_addr_ = nullptr;
+  size_t io_addr_size_ = 0;
+
+  // host addr
+  std::vector<void *> io_addr_host_;
 
   bool dynamic_flag_ = false;
   // for copy task
-  void *copy_task_args_buf_;
-  void *copy_workspace_buf_;
+  void *copy_task_args_buf_ = nullptr;
+  void *copy_workspace_buf_ = nullptr;
 
   std::vector<void *> output_summary_;
   std::vector<aicpu::FWKAdapter::ResultSummary> output_summary_host_;
 
-  void *copy_ioaddr_dev_;
+  void *copy_ioaddr_dev_ = nullptr;
 
-  void *copy_input_release_flag_dev_;
-  void *copy_input_data_size_dev_;
-  void *copy_input_src_dev_;
-  void *copy_input_dst_dev_;
+  void *copy_input_release_flag_dev_ = nullptr;
+  void *copy_input_data_size_dev_ = nullptr;
+  void *copy_input_src_dev_ = nullptr;
+  void *copy_input_dst_dev_ = nullptr;
 
   vector<void *> out_shape_hbm_;
   uint64_t kernel_id_ = 0;
@@ -222,13 +200,12 @@ class AiCpuCCTask : public AiCpuBaseTask {
   AiCpuCCTask &operator=(const AiCpuCCTask &) = delete;
 
   Status LaunchKernel(rtStream_t stream) override;
-  OpTaskType GetOpTaskType() override { return OP_TASK_AICPUCC; }
-  const void *GetIOAddr() const override;
+  void GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) override;
   const void *GetArgs() const;
   void SetKernelArgs(std::unique_ptr<uint8_t[]> args, size_t arg_size);
   void SetSoName(const std::string &so_name);
   void SetkernelName(const std::string &kernel_Name);
-  void SetIoAddr(void *io_addr);
+  void SetIoAddr(uintptr_t *io_addr);
   size_t GetArgSize() const;
 
   Status LaunchKernel(const std::vector<GeTensorDesc> &input_desc,
@@ -244,7 +221,8 @@ private:
   std::unique_ptr<uint8_t[]> args_;
   size_t arg_size_ = 0;
   void *sm_desc_ = nullptr;
-  void *io_addr_ = nullptr;
+  uintptr_t *io_addr_ = nullptr;
+  size_t io_addr_num_ = 0;
   bool is_custom_ = false;
   uint32_t dump_flag_ = RT_KERNEL_DEFAULT;
 };