Optimize Task sink memory usage.

ge/graph/load/new_model_manager/davinci_model.cc

@@ -3021,7 +3021,14 @@ Status DavinciModel::MallocKnownArgs() {
     GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
     return RT_ERROR_TO_GE_STATUS(rt_ret);
   }
-
+  // malloc dynamic and static hybrid memory
+  if (total_hybrid_args_size_ != 0) {
+    rt_ret = rtMalloc(&hybrid_addrs_, total_hybrid_args_size_, RT_MEMORY_HBM);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
+      return RT_ERROR_TO_GE_STATUS(rt_ret);
+    }
+  }
   // malloc fixed addr memory, eg: rts op
   if (total_fixed_addr_size_ != 0) {
     GELOGI("Begin to allocate fixed addr.");

ge/graph/load/new_model_manager/davinci_model.h

@@ -498,6 +498,14 @@ class DavinciModel {
     return cur_args;
   }
   void SetTotalIOAddrs(const vector<void *> &io_addrs);
+  void SetHybridArgsSize(uint32_t args_size) { total_hybrid_args_size_ += args_size; }
+  uint32_t GetHybridArgsSize() {
+    return total_hybrid_args_size_;
+  }
+  void *GetCurrentHybridArgsAddr(uint32_t offset) {
+    void *cur_args = static_cast<char *>(hybrid_addrs_) + offset;
+    return cur_args;
+  }
   void SetTotalFixedAddrsSize(string tensor_name, int64_t fix_addr_size);
   int64_t GetFixedAddrsSize(string tensor_name);
   void *GetCurrentFixedAddr(int64_t offset) const {
@@ -975,6 +983,8 @@ class DavinciModel {
   void *args_ = nullptr;
   void *args_host_ = nullptr;
   void *fixed_addrs_ = nullptr;
+  void *hybrid_addrs_ = nullptr;
+  uint32_t total_hybrid_args_size_ = 0;
   int64_t total_fixed_addr_size_ = 0;
   map<const void *, void *> known_input_data_info_;
   map<const void *, void *> known_output_data_info_;

ge/graph/load/new_model_manager/task_info/kernel_task_info.cc

@@ -444,7 +444,6 @@ void KernelTaskInfo::SetIoAddrs(const OpDescPtr &op_desc) {
     vector<void *> workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc);
     io_addrs_.insert(io_addrs_.end(), workspace_data_addrs.begin(), workspace_data_addrs.end());
   }
-  return SUCCESS;
 }
 
 Status KernelTaskInfo::UpdateArgs() {
@@ -897,6 +896,12 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k
 
   aicpu_param_head->extInfoAddr = reinterpret_cast<uintptr_t>(aicpu_ext_info_addr_);
   aicpu_param_head->extInfoLength = static_cast<uintptr_t>(ext_info.size());
+
+  if (davinci_model_->IsKnownNode()) {
+    args_ = davinci_model_->GetCurrentHybridArgsAddr(hybrid_args_offset_);
+    InitDumpTask(sizeof(aicpu::AicpuParamHead));
+    return SUCCESS;
+  }
   const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam();
   vector<void *> input_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc);
   vector<void *> output_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc);