Optional output does not allocate memory

ge/graph/build/memory/block_mem_assigner.cc

@@ -24,6 +24,7 @@
 #include "graph/buffer.h"
 #include "graph/ge_attr_value.h"
 #include "graph/ge_context.h"
+#include "graph/types.h"
 #include "graph/node.h"
 #include "graph/utils/graph_utils.h"
 #include "graph/utils/node_utils.h"
@@ -1401,6 +1402,7 @@ Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector
     if (output_op_desc != nullptr) {
       GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS, GELOGI("Get size failed"));
     }
+
     // fusion: other type's size not means malloc HBM memory
     bool l1_flag = has_mem_type_attr && memorys_type[i] == RT_MEMORY_L1;
     if (l1_flag) {
@@ -1408,6 +1410,11 @@ Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector
              op_desc->GetName().c_str(), op_desc->GetOutputNameByIndex(i).c_str(), memorys_type[i]);
       size = 0;
     }
+
+    int32_t calc_type = 0;
+    bool ret = ge::AttrUtils::GetInt(output_op_desc, ATTR_NAME_MEMORY_SIZE_CALC_TYPE, calc_type);
+    GE_IF_BOOL_EXEC((ret && (calc_type == static_cast<int32_t>(ge::MemorySizeCalcType::ALWAYS_EMPTY))), size = 0;);
+
     std::string peer_name;
     uint32_t peer_input_index = 0;
     bool out_node_set_continuous_input = false;

ge/graph/load/new_model_manager/model_utils.cc

@@ -20,6 +20,7 @@
 #include "common/op/ge_op_utils.h"
 #include "graph/utils/tensor_utils.h"
 #include "graph/manager/graph_var_manager.h"
+#include "graph/types.h"
 
 #define VALIDATE_MEM_RANGE(OP, SIZE, OFFSET)                                                                 \
   do {                                                                                                       \
@@ -340,7 +341,7 @@ vector<void *> ModelUtils::GetInputDataAddrs(const RuntimeParam &model_param, Co
                     GELOGI("[IMAS]GetInputDataAddrs graph_%u type[V] name[%s] input[%lu] memaddr[%p]",
                            model_param.graph_id, op_desc->GetName().c_str(), i, variable_addr);
                     continue);
-    
+
     int64_t mem_type;
     bool tensor_has_mem_type = ge::AttrUtils::GetInt(tensor_desc, ATTR_NAME_TENSOR_MEM_TYPE, mem_type);
     // feature maps
@@ -424,6 +425,18 @@ vector<void *> ModelUtils::GetOutputDataAddrs(const RuntimeParam &model_param, C
     return v_output_data_addr;
   }
   for (size_t i = 0; i < outputs_size; ++i) {
+    const GeTensorDescPtr tensor_desc = op_desc->MutableOutputDesc(i);
+    if (tensor_desc == nullptr) {
+      GELOGW("Op: %s, Index: %zu, Tensor Desc is null", op_desc->GetName().c_str(), i);
+      continue;
+    }
+
+    int32_t calc_type = 0;
+    bool ret = ge::AttrUtils::GetInt(tensor_desc, ATTR_NAME_MEMORY_SIZE_CALC_TYPE, calc_type);
+    if (ret && (calc_type == static_cast<int32_t>(ge::MemorySizeCalcType::ALWAYS_EMPTY))) {
+      GELOGD("%s is an optional output, the address don't need to be saved.", tensor_desc->GetName().c_str());
+      continue;
+    }
     GE_IF_BOOL_EXEC(model_param.var_size != 0 && ge::VarManager::Instance(session_id)->IsVarAddr(v_output_offset[i]),
                     uint8_t *variable_addr = nullptr;
                     GE_CHK_STATUS_EXEC(GetVarAddr(model_param, op_desc, v_output_offset[i], variable_addr), return {});
@@ -431,11 +444,6 @@ vector<void *> ModelUtils::GetOutputDataAddrs(const RuntimeParam &model_param, C
                     GELOGI("[IMAS]GetOutputDataAddrs graph_%u type[V] name[%s] output[%zu] memaddr[%p]",
                            model_param.graph_id, op_desc->GetName().c_str(), i, variable_addr);
                     continue);
-    const GeTensorDescPtr tensor_desc = op_desc->MutableOutputDesc(i);
-    if (tensor_desc == nullptr) {
-      GELOGW("Op: %s, Index: %zu, Tensor Desc is null", op_desc->GetName().c_str(), i);
-      continue;
-    }
 
     int64_t mem_type;
     bool tensor_has_mem_type = ge::AttrUtils::GetInt(tensor_desc, ATTR_NAME_TENSOR_MEM_TYPE, mem_type);

ge/hybrid/node_executor/aicore/aicore_op_task.cc

@@ -20,6 +20,7 @@
 #include "hybrid/executor/hybrid_execution_context.h"
 #include "hybrid/node_executor/aicore/aicore_task_builder.h"
 #include "graph/load/new_model_manager/tbe_handle_store.h"
+#include "graph/types.h"
 
 using optiling::OpRunInfo;
 
@@ -34,6 +35,23 @@ constexpr char const *kAttrAtomicOpParamSize = "atomic_op_para_size";
 Status AiCoreOpTask::Init(const OpDesc &op_desc, const domi::TaskDef &task_def) {
   GE_CHK_STATUS_RET_NOLOG(InitWithTaskDef(op_desc, task_def));
   GE_CHK_STATUS_RET_NOLOG(InitTilingInfo(op_desc));
+
+  GE_CHECK_LE(op_desc.GetOutputsSize(), static_cast<size_t>(INT_MAX));
+  int outputs_size = static_cast<int>(op_desc.GetOutputsSize());
+
+  for (int i = 0; i < outputs_size; ++i) {
+    const GeTensorDescPtr tensor_desc = op_desc.MutableOutputDesc(i);
+    if (tensor_desc == nullptr) {
+      GELOGW("Op: %s, Index: %d, Tensor Desc is null", op_desc.GetName().c_str(), i);
+      continue;
+    }
+
+    int32_t calc_type = 0;
+    bool ret = ge::AttrUtils::GetInt(tensor_desc, ATTR_NAME_MEMORY_SIZE_CALC_TYPE, calc_type);
+    if (ret && (calc_type == static_cast<int32_t>(ge::MemorySizeCalcType::ALWAYS_EMPTY))) {
+      output_indices_to_skip_.push_back(i);
+    }
+  }
   return SUCCESS;
 }
 
@@ -221,7 +239,8 @@ Status AiCoreOpTask::CalcTilingInfo(const NodePtr &node, OpRunInfo &tiling_info)
 }
 
 Status AiCoreOpTask::UpdateArgs(TaskContext &task_context) {
-  size_t expected_arg_count = task_context.NumInputs() + task_context.NumOutputs() + task_context.NumWorkspaces();
+  size_t expected_arg_count = task_context.NumInputs() + task_context.NumOutputs() + task_context.NumWorkspaces()
+                              - output_indices_to_skip_.size();
   if (tiling_buffer_ != nullptr) {
     ++expected_arg_count;
   }
@@ -244,6 +263,11 @@ Status AiCoreOpTask::UpdateArgs(TaskContext &task_context) {
   for (int i = 0; i < task_context.NumOutputs(); ++i) {
     const auto output = task_context.GetOutput(i);
     GE_CHECK_NOTNULL(output);
+    if (find(output_indices_to_skip_.begin(), output_indices_to_skip_.end(), i) != output_indices_to_skip_.end()) {
+      GELOGD("Node:%s output[%d] is an optional, the address don't need to be saved.",
+             task_context.GetNodeName(), i);
+      continue;
+    }
     arg_base_[index++] = reinterpret_cast<uintptr_t>(output->GetData());
   }
 

ge/hybrid/node_executor/aicore/aicore_op_task.h

@@ -72,6 +72,7 @@ class AiCoreOpTask {
   uint32_t args_size_ = 0;
   uint32_t block_dim_ = 1;
   bool clear_atomic_ = true;
+  std::vector<int> output_indices_to_skip_;
 };
 
 class AtomicAddrCleanOpTask : public AiCoreOpTask {

ge/hybrid/node_executor/task_context.cc

@@ -18,6 +18,7 @@
 #include "framework/common/ge_inner_error_codes.h"
 #include "framework/common/debug/log.h"
 #include "graph/utils/tensor_utils.h"
+#include "graph/types.h"
 #include "graph/debug/ge_attr_define.h"
 #include "hybrid/executor/hybrid_execution_context.h"
 #include "hybrid/executor/subgraph_executor.h"
@@ -213,6 +214,13 @@ Status TaskContext::AllocateOutput(int index,
     return SUCCESS;
   }
 
+  int32_t calc_type = 0;
+  bool ret = ge::AttrUtils::GetInt(tensor_desc, ATTR_NAME_MEMORY_SIZE_CALC_TYPE, calc_type);
+  if (ret && (calc_type == static_cast<int32_t>(ge::MemorySizeCalcType::ALWAYS_EMPTY))) {
+    outputs_start_[index] = TensorValue();
+    return SUCCESS;
+  }
+
   auto it = node_item_->ref_outputs.find(index);
   if (it != node_item_->ref_outputs.end()) {
     auto &ref_node = it->second;

metadef

@@ -1 +1 @@
-Subproject commit b00c50c2a8c2ce06929b27f7b74185a950737ec8
+Subproject commit 88d053a5f94c40ff21620cef50b87075d5054292

parser

@@ -1 +1 @@
-Subproject commit f0109a2c70981d74932bb38bb56722caff3323a5
+Subproject commit 6904ba9488658afc30076d299183fc8875045f49