!2028 Fix bug of single_op.

Merge pull request !2028 from 赵之轩/my_dev3

ge/single_op/single_op.cc

@@ -433,11 +433,13 @@ Status DynamicSingleOp::ExecuteAsync(const vector<GeTensorDesc> &input_desc,
   if (!inputs_size.empty()) {
     StreamResource *stream_resource  = SingleOpManager::GetInstance().GetResource(resource_id_, stream_);
     GE_CHK_STATUS_RET_NOLOG(UpdateInputsBufferAddr(stream_resource, stream_, inputs_size, update_buffers));
-    GE_CHK_STATUS_RET_NOLOG(SetHostTensorValue(input_desc, input_buffers));
   }
 
   if (hybrid_model_executor_ != nullptr) {
     GELOGD("Execute multi-task dynamic single op by hybrid model executor");
+    if (!inputs_size.empty()) {
+      GE_CHK_STATUS_RET_NOLOG(SetHostTensorValue(input_desc, input_buffers));
+    }
     hybrid::HybridModelExecutor::ExecuteArgs args;
     GE_CHK_STATUS_RET_NOLOG(InitHybridModelArgs(update_buffers, output_buffers, input_desc, args));
 

ge/single_op/task/op_task.cc

@@ -293,6 +293,9 @@ Status TbeOpTask::UpdateNodeByShape(const vector<GeTensorDesc> &input_desc, cons
 }
 
 Status TbeOpTask::EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size) {
+  node_ = node;
+  tiling_buffer_ = tiling_buffer;
+  max_tiling_size_ = max_tiling_size;
   if (tiling_buffer != nullptr) {
     uintptr_t *arg_base = nullptr;
     size_t arg_num = 0;
@@ -310,9 +313,6 @@ Status TbeOpTask::EnableDynamicSupport(const NodePtr &node, void *tiling_buffer,
     }
     arg_base[tiling_index] = reinterpret_cast<uintptr_t>(tiling_buffer);
   }
-  node_ = node;
-  tiling_buffer_ = tiling_buffer;
-  max_tiling_size_ = max_tiling_size;
   return SUCCESS;
 }
 
@@ -481,6 +481,25 @@ void TbeOpTask::GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) {
   }
 }
 
+Status AtomicAddrCleanOpTask::EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size) {
+  node_ = node;
+  tiling_buffer_ = tiling_buffer;
+  max_tiling_size_ = max_tiling_size;
+  if (tiling_buffer != nullptr) {
+    uintptr_t *arg_base = nullptr;
+    size_t arg_num = 0;
+    GetIoAddr(arg_base, arg_num);
+    uint32_t tiling_index = atomic_output_indices_.size();
+    if (arg_num == 0 || arg_num < tiling_index) {
+      GELOGE(ACL_ERROR_GE_INTERNAL_ERROR, "[Check][Size]Tiling index %u, arg number %zu is invalid.",
+             tiling_index, arg_num);
+      return ACL_ERROR_GE_INTERNAL_ERROR;
+    }
+    arg_base[tiling_index] = reinterpret_cast<uintptr_t>(tiling_buffer);
+  }
+  return SUCCESS;
+}
+
 Status AtomicAddrCleanOpTask::UpdateNodeByShape(const vector<GeTensorDesc> &input_desc,
                                                 const vector<GeTensorDesc> &output_desc) {
   return SUCCESS;

ge/single_op/task/op_task.h

@@ -97,7 +97,7 @@ class TbeOpTask : public OpTask {
   const void *GetArgs() const;
   size_t GetArgSize() const;
   const std::string &GetStubName() const;
-  Status EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size);
+  virtual Status EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size);
   const std::string &GetTaskType() const override;
   void SetHandle(void *handle);
 
@@ -149,6 +149,7 @@ class TbeOpTask : public OpTask {
 class AtomicAddrCleanOpTask : public TbeOpTask {
  public:
   Status InitAtomicAddrCleanIndices();
+  Status EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size) override;
 
  private:
   Status UpdateNodeByShape(const vector<GeTensorDesc> &input_desc,
@@ -156,8 +157,8 @@ class AtomicAddrCleanOpTask : public TbeOpTask {
   Status UpdateIoAddr(const vector<DataBuffer> &inputs, const vector<DataBuffer> &outputs) override;
   Status UpdateTilingArgs(rtStream_t stream) override;
   Status CalcTilingInfo(optiling::utils::OpRunInfo &run_info) override;
-  std::vector<int> atomic_output_indices_;
 
+  std::vector<int> atomic_output_indices_;
 };
 
 class AiCpuBaseTask : public OpTask {

ge/single_op/task/tbe_task_builder.cc

@@ -425,7 +425,7 @@ Status TbeTaskBuilder::InitTilingInfo(TbeOpTask &task) {
     GELOGD("[%s] Done allocating tiling buffer, size=%ld.", op_desc_->GetName().c_str(), max_size);
   }
 
-  task.EnableDynamicSupport(node_, tiling_buffer, static_cast<uint32_t>(max_size));
+  GE_CHK_STATUS_RET_NOLOG(task.EnableDynamicSupport(node_, tiling_buffer, static_cast<uint32_t>(max_size)));
   return SUCCESS;
 }
 

tests/ut/ge/single_op/single_op_task_unittest.cc

@@ -237,3 +237,27 @@ TEST_F(UtestSingleOpTask, test_aicpu_task_update_io_addr) {
     ASSERT_EQ(ret, PARAM_INVALID);
   }
 }
+
+TEST_F(UtestSingleOpTask, test_dynamic_support) {
+  auto graph = make_shared<ComputeGraph>("graph");
+  auto op_desc = make_shared<OpDesc>("Add", "Add");
+  auto node = graph->AddNode(op_desc);
+  AtomicAddrCleanOpTask atomic_task;
+  TbeOpTask tbe_task;
+
+  tbe_task.arg_size_ = sizeof(void *) * 1;
+  tbe_task.args_.reset(new (std::nothrow) uint8_t[tbe_task.arg_size_]);
+  atomic_task.arg_size_ = sizeof(void *) * 1;
+  atomic_task.args_.reset(new (std::nothrow) uint8_t[atomic_task.arg_size_]);
+  ASSERT_EQ(tbe_task.EnableDynamicSupport(node, (void *)0x0001, 1), ACL_ERROR_GE_INTERNAL_ERROR);
+  ASSERT_EQ(atomic_task.EnableDynamicSupport(node, (void *)0x0001, 1), ACL_ERROR_GE_INTERNAL_ERROR);
+
+  tbe_task.arg_size_ = sizeof(void *) * 2;
+  tbe_task.args_.reset(new (std::nothrow) uint8_t[tbe_task.arg_size_]);
+  atomic_task.arg_size_ = sizeof(void *) * 2;
+  atomic_task.args_.reset(new (std::nothrow) uint8_t[atomic_task.arg_size_]);
+  ASSERT_EQ(tbe_task.EnableDynamicSupport(node, (void *)0x0001, 1), SUCCESS);
+  ASSERT_EQ(atomic_task.EnableDynamicSupport(node, (void *)0x0001, 1), SUCCESS);
+  tbe_task.tiling_buffer_ = nullptr;
+  atomic_task.tiling_buffer_ = nullptr;
+}