!1445 Save atomic kernel bin to model.

From: @zhao_zhixuan
Reviewed-by: @xchu42,@ji_chen
Signed-off-by: @ji_chen

ge/common/tbe_kernel_store.cc

@@ -15,6 +15,8 @@
  */
 
 #include "common/tbe_kernel_store.h"
+#include "graph/utils/attr_utils.h"
+#include "graph/debug/ge_attr_define.h"
 
 namespace ge {
 
@@ -31,6 +33,15 @@ void TBEKernelStore::LoadTBEKernelBinToOpDesc(const std::shared_ptr<ge::OpDesc>
       GE_IF_BOOL_EXEC(!op_desc->SetExtAttr(ge::OP_EXTATTR_NAME_TBE_KERNEL, kernel_bin),
                       GELOGW("LoadKernelTBEBinToOpDesc: SetExtAttr for kernel_bin failed");)
       GELOGI("Load tbe kernel:%s, %zu", kernel_bin->GetName().c_str(), kernel_bin->GetBinDataSize());
+
+      std::string atomic_kernel_name;
+      (void) AttrUtils::GetStr(op_desc, ATOMIC_ATTR_TBE_KERNEL_NAME, atomic_kernel_name);
+      if (!atomic_kernel_name.empty()) {
+        GELOGI("Get atomic kernel name is %s.", atomic_kernel_name.c_str());
+        auto atomic_kernel_bin = FindKernel(atomic_kernel_name);
+        GE_IF_BOOL_EXEC(!op_desc->SetExtAttr(EXT_ATTR_ATOMIC_TBE_KERNEL, atomic_kernel_bin),
+                        GELOGW("LoadKernelTBEBinToOpDesc: SetExtAttr for atomic kernel_bin failed");)
+      }
     }
   }
 }

ge/graph/build/model_builder.cc

@@ -627,6 +627,50 @@ Status ModelBuilder::MergeWeights() {
   return SUCCESS;
 }
 
+Status ModelBuilder::SaveAtomicTBEKernel(const OpDescPtr &op_desc) {
+  ge::NodePtr atomic_clean_node = nullptr;
+  atomic_clean_node = op_desc->TryGetExtAttr("atomic_clean_node_ptr", atomic_clean_node);
+  if (atomic_clean_node == nullptr) {
+    return SUCCESS;
+  }
+
+  ge::OpDescPtr atomic_op_desc = atomic_clean_node->GetOpDesc();
+  GE_CHECK_NOTNULL(atomic_op_desc);
+  TBEKernelPtr tbe_kernel = atomic_op_desc->TryGetExtAttr(ge::OP_EXTATTR_NAME_TBE_KERNEL, TBEKernelPtr());
+  if (tbe_kernel == nullptr) {
+    std::string kernel_name;
+    GeAttrValue::BYTES kernel_buffer;
+    (void) AttrUtils::GetStr(atomic_op_desc, ATTR_NAME_TBE_KERNEL_NAME, kernel_name);
+    (void) AttrUtils::GetBytes(atomic_op_desc, ATTR_NAME_TBE_KERNEL_BUFFER, kernel_buffer);
+    if (!kernel_name.empty() && (kernel_buffer.GetSize() > 0)) {
+      GE_CHECK_NOTNULL(kernel_buffer.GetData());
+      std::vector<char> data(kernel_buffer.GetData(), kernel_buffer.GetData() + kernel_buffer.GetSize());
+      tbe_kernel = MakeShared<OpKernelBin>(kernel_name, std::move(data));
+      GE_CHECK_NOTNULL(tbe_kernel);
+    }
+  }
+  if (tbe_kernel == nullptr) {
+    GELOGD("Atomic_clean_node doesn't have tbe_kernel.");
+    return SUCCESS;
+  }
+  tbe_kernel_store_.AddTBEKernel(tbe_kernel);
+  GELOGD("Atomic_clean_node tbe_kernel_name %s!", tbe_kernel->GetName().c_str());
+  (void) AttrUtils::SetStr(op_desc, ATOMIC_ATTR_TBE_KERNEL_NAME, tbe_kernel->GetName());
+
+  std::string kernel_name;
+  (void) AttrUtils::GetStr(atomic_op_desc, atomic_op_desc->GetName() + "_kernelname", kernel_name);
+  (void) AttrUtils::SetStr(op_desc, op_desc->GetName() + "_atomic_kernelname", kernel_name);
+
+  std::string meta_data;
+  (void) AttrUtils::GetStr(atomic_op_desc, TVM_ATTR_NAME_METADATA, meta_data);
+  (void) AttrUtils::SetStr(op_desc, ATOMIC_ATTR_TVM_METADATA, meta_data);
+
+  std::string json_string;
+  (void) AttrUtils::GetStr(atomic_op_desc, TVM_ATTR_NAME_MAGIC, json_string);
+  (void) AttrUtils::SetStr(op_desc, ATOMIC_ATTR_TVM_MAGIC, json_string);
+  return SUCCESS;
+}
+
 Status ModelBuilder::SaveDataToModel(ge::Model &model, ge::GeModel &ge_model) {
   // Add weight
   ge_model.SetWeight(weight_buffer_);
@@ -662,6 +706,8 @@ Status ModelBuilder::SaveDataToModel(ge::Model &model, ge::GeModel &ge_model) {
     }
     tbe_name_set.insert(tbe_kernel->GetName());
     tbe_kernel_store_.AddTBEKernel(tbe_kernel);
+
+    GE_CHK_STATUS_RET(SaveAtomicTBEKernel(node_op_desc), "[Save][TBEKernel] save atomic tbekernel failed!");
   }
 
   SetModelCheckAicpuAttr(model, aicpu_op_types, aicpu_tf_op_types);

ge/graph/build/model_builder.h

@@ -89,6 +89,8 @@ class ModelBuilder {
   void SetModelCheckAicpuAttr(ge::Model &model, std::set<std::string> &aicpu_op_types,
                                 std::set<std::string> &aicpu_tf_op_types);
 
+  Status SaveAtomicTBEKernel(const OpDescPtr &op_desc);
+
   uint64_t session_id_;
 
   map<int64_t, size_t> mem_type_to_mem_offset_;

ge/hybrid/node_executor/aicore/aicore_op_task.cc

@@ -71,22 +71,22 @@ Status AiCoreOpTask::Init(const OpDesc &op_desc, const domi::TaskDef &task_def)
 }
 
 Status AiCoreOpTask::RegisterTbeHandle(const OpDesc &op_desc) {
-  auto op_desc_ptr = std::make_shared<OpDesc>(op_desc);
-  GE_CHECK_NOTNULL(op_desc_ptr);
-  auto tbe_kernel = op_desc_ptr->TryGetExtAttr(OP_EXTATTR_NAME_TBE_KERNEL, TBEKernelPtr());
-  if (tbe_kernel == nullptr) {
-    GELOGE(INTERNAL_ERROR, "TBE: %s can't find tvm bin file!", op_desc_ptr->GetName().c_str());
-    return INTERNAL_ERROR;
-  }
-  TBEHandleStore &kernel_store = TBEHandleStore::GetInstance();
   rtError_t rt_ret = rtQueryFunctionRegistered(stub_name_.c_str());
   if (rt_ret != RT_ERROR_NONE || is_single_op_) {
+    auto op_desc_ptr = MakeShared<OpDesc>(op_desc);
+    GE_CHECK_NOTNULL(op_desc_ptr);
+    auto tbe_kernel = op_desc_ptr->TryGetExtAttr(GetKeyForTbeKernel(), TBEKernelPtr());
+    if (tbe_kernel == nullptr) {
+      GELOGE(INTERNAL_ERROR, "TBE: %s can't find tvm bin file!", op_desc_ptr->GetName().c_str());
+      return INTERNAL_ERROR;
+    }
+    TBEHandleStore &kernel_store = TBEHandleStore::GetInstance();
     void *bin_handle = nullptr;
     if (!kernel_store.FindTBEHandle(stub_name_.c_str(), bin_handle)) {
       GELOGI("TBE: can't find the binfile_key[%s] in HandleMap", stub_name_.c_str());
       rtDevBinary_t binary;
       std::string json_string;
-      GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc_ptr, TVM_ATTR_NAME_MAGIC, json_string),
+      GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc_ptr, GetKeyForTvmMagic(), json_string),
                       GELOGI("Get original type of session_graph_id."));
       if (json_string == "RT_DEV_BINARY_MAGIC_ELF_AICPU") {
         binary.magic = RT_DEV_BINARY_MAGIC_ELF_AICPU;
@@ -104,7 +104,7 @@ Status AiCoreOpTask::RegisterTbeHandle(const OpDesc &op_desc) {
       GELOGI("TBE: binary.length: %lu", binary.length);
       GE_CHK_RT_RET(rtDevBinaryRegister(&binary, &bin_handle));
       std::string meta_data;
-      GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc_ptr, TVM_ATTR_NAME_METADATA, meta_data),
+      GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc_ptr, GetKeyForTvmMetaData(), meta_data),
                       GELOGI("Get original type of json_string"));
       GELOGI("TBE: meta data: %s", meta_data.empty() ? "null" : meta_data.c_str());
       GE_IF_BOOL_EXEC(!meta_data.empty(), GE_CHK_RT_RET(rtMetadataRegister(bin_handle, meta_data.c_str())));
@@ -114,7 +114,7 @@ Status AiCoreOpTask::RegisterTbeHandle(const OpDesc &op_desc) {
       kernel_store.ReferTBEHandle(stub_name_.c_str());
     }
     std::string kernel_name;
-    GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc_ptr, op_desc_ptr->GetName() + "_kernelname", kernel_name),
+    GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc_ptr, GetKeyForKernelName(op_desc), kernel_name),
                     GELOGI("Get original type of kernel_name"));
     GELOGI("TBE: binfile_key=%s, kernel_name=%s", stub_name_.c_str(), kernel_name.c_str());
     GE_CHK_RT_RET(rtFunctionRegister(bin_handle, stub_name_.c_str(), stub_name_.c_str(), kernel_name.c_str(), 0));
@@ -349,9 +349,6 @@ Status AiCoreOpTask::CalcTilingInfo(const NodePtr &node, OpRunInfo &tiling_info)
   GE_CHK_STATUS_RET(OpParaCalculate(*node, tiling_info),
                     "Failed calc tiling data of node %s.",
                     node->GetName().c_str());
-  if (is_single_op_) {
-    tiling_info.clear_atomic = false;
-  }
   GELOGD("[%s] Done invoking OpParaCalculate successfully.", node->GetName().c_str());
   return SUCCESS;
 }
@@ -468,6 +465,22 @@ std::string AiCoreOpTask::GetKeyForOpParamSize() const {
   return kAttrOpParamSize;
 }
 
+std::string AiCoreOpTask::GetKeyForTbeKernel() const {
+  return OP_EXTATTR_NAME_TBE_KERNEL;
+}
+
+std::string AiCoreOpTask::GetKeyForTvmMagic() const {
+  return TVM_ATTR_NAME_MAGIC;
+}
+
+std::string AiCoreOpTask::GetKeyForTvmMetaData() const {
+  return TVM_ATTR_NAME_METADATA;
+}
+
+std::string AiCoreOpTask::GetKeyForKernelName(const OpDesc &op_desc) const {
+  return op_desc.GetName() + "_kernelname";
+}
+
 Status AtomicAddrCleanOpTask::Init(const OpDesc &op_desc, const domi::TaskDef &task_def) {
   GE_CHK_STATUS_RET_NOLOG(AiCoreOpTask::Init(op_desc, task_def));
   return InitAtomicAddrCleanIndices(op_desc);
@@ -524,6 +537,22 @@ std::string AtomicAddrCleanOpTask::GetKeyForOpParamSize() const {
   return kAttrAtomicOpParamSize;
 }
 
+std::string AtomicAddrCleanOpTask::GetKeyForTbeKernel() const {
+  return EXT_ATTR_ATOMIC_TBE_KERNEL;
+}
+
+std::string AtomicAddrCleanOpTask::GetKeyForTvmMagic() const {
+  return ATOMIC_ATTR_TVM_MAGIC;
+}
+
+std::string AtomicAddrCleanOpTask::GetKeyForTvmMetaData() const {
+  return ATOMIC_ATTR_TVM_METADATA;
+}
+
+std::string AtomicAddrCleanOpTask::GetKeyForKernelName(const OpDesc &op_desc) const {
+  return op_desc.GetName() + "_atomic_kernelname";
+}
+
 Status AtomicAddrCleanOpTask::CalcTilingInfo(const NodePtr &node, OpRunInfo &tiling_info) {
   GELOGD("[%s] Start to invoke OpAtomicCalculate.", node->GetName().c_str());
   GE_CHK_STATUS_RET(OpAtomicCalculate(*node, tiling_info),

ge/hybrid/node_executor/aicore/aicore_op_task.h

@@ -81,6 +81,10 @@ class AiCoreOpTask {
  protected:
   Status UpdateTilingInfo(TaskContext &context);
   virtual std::string GetKeyForOpParamSize() const;
+  virtual std::string GetKeyForTbeKernel() const;
+  virtual std::string GetKeyForTvmMagic() const;
+  virtual std::string GetKeyForTvmMetaData() const;
+  virtual std::string GetKeyForKernelName(const OpDesc &op_desc) const;
   virtual Status CalcTilingInfo(const NodePtr &node, optiling::OpRunInfo &tiling_info);
 
   std::unique_ptr<TensorBuffer> tiling_buffer_ = nullptr;
@@ -119,6 +123,10 @@ class AtomicAddrCleanOpTask : public AiCoreOpTask {
 
  protected:
   std::string GetKeyForOpParamSize() const override;
+  std::string GetKeyForTbeKernel() const override;
+  std::string GetKeyForTvmMagic() const override;
+  std::string GetKeyForTvmMetaData() const override;
+  std::string GetKeyForKernelName(const OpDesc &op_desc) const override;
   Status CalcTilingInfo(const NodePtr &node, optiling::OpRunInfo &tiling_info) override;
 
  private:

ge/hybrid/node_executor/aicore/aicore_task_builder.cc

@@ -70,6 +70,7 @@ Status AiCoreTaskBuilder::BuildTask(std::unique_ptr<AiCoreNodeTask> &node_task,
     auto atomic_task =
         std::unique_ptr<AtomicAddrCleanOpTask>(new(std::nothrow)AtomicAddrCleanOpTask());
     GE_CHECK_NOTNULL(atomic_task);
+    atomic_task->SetSingleOp(is_single_op);
     GE_CHK_STATUS_RET(atomic_task->Init(*op_desc_, task_defs_.front()),
                       "[%s] Failed to init task for AtomicAddrClean",
                       op_desc_->GetName().c_str());

metadef

@@ -1 +1 @@
-Subproject commit 4ff5e3987f2e5d2980019defacaf0891861c84fc
+Subproject commit 366b15574218befa11454311879a4f436eeb67a9

parser

@@ -1 +1 @@
-Subproject commit 51fb6c4850906e8342598d47eccfca0b87ffea59
+Subproject commit d744541c6ca7f6966c1befacc9f83f53b0829e0a

tests/ut/ge/graph/build/model_builder_unittest.cc

@@ -144,3 +144,20 @@ TEST_F(UtestModelBuilderTest, SetInputIsConst) {
   ge::ModelBuilder builder(0, graph, subgraphs, stream_max_parallel_num, false);
   EXPECT_EQ(builder.PreBuildModel(), SUCCESS);
 }
+
+TEST_F(UtestModelBuilderTest, test_save_atomic_bin) {
+  Graph2SubGraphInfoList subgraphs;
+  std::map<std::string, int> stream_max_parallel_num;
+  ge::ComputeGraphPtr graph = make_shared<ge::ComputeGraph>("");
+  ge::ModelBuilder builder(0, graph, subgraphs, stream_max_parallel_num, false);
+
+  auto atomic_op_desc = make_shared<OpDesc>("Atomic", "Atomic");
+  auto kernel_buffer = static_cast<GeAttrValue::BYTES>(Buffer(10));
+  AttrUtils::SetStr(atomic_op_desc, ATTR_NAME_TBE_KERNEL_NAME, "Atomic");
+  AttrUtils::SetBytes(atomic_op_desc, ATTR_NAME_TBE_KERNEL_BUFFER, kernel_buffer);
+
+  ge::NodePtr atomic_node = graph->AddNode(atomic_op_desc);
+  auto op_desc = make_shared<OpDesc>("Sum", "Sum");
+  op_desc->SetExtAttr("atomic_clean_node_ptr", atomic_node);
+  EXPECT_EQ(builder.SaveAtomicTBEKernel(op_desc), SUCCESS);
+}

tests/ut/ge/hybrid/ge_hybrid_unittest.cc

@@ -540,3 +540,18 @@ TEST_F(UtestGeHybrid, TestOptimizeDependenciesForConstInputs) {
   ASSERT_EQ(dst_node_item->dependents_for_shape_inference.size(), 1);
   ASSERT_EQ(dst_node_item->dependents_for_shape_inference[0], non_const_node_item->node);
 }
+
+TEST_F(UtestGeHybrid, test_key_for_kernel_bin) {
+  auto aicore_task = std::unique_ptr<hybrid::AiCoreOpTask>(new(std::nothrow)hybrid::AiCoreOpTask());
+  OpDesc op_desc("Sum", "Sum");
+  EXPECT_EQ(aicore_task->GetKeyForTbeKernel(), OP_EXTATTR_NAME_TBE_KERNEL);
+  EXPECT_EQ(aicore_task->GetKeyForTvmMagic(), TVM_ATTR_NAME_MAGIC);
+  EXPECT_EQ(aicore_task->GetKeyForTvmMetaData(), TVM_ATTR_NAME_METADATA);
+  EXPECT_EQ(aicore_task->GetKeyForKernelName(op_desc), "Sum_kernelname");
+
+  auto atomic_task = std::unique_ptr<hybrid::AtomicAddrCleanOpTask>(new(std::nothrow)hybrid::AtomicAddrCleanOpTask());
+  EXPECT_EQ(atomic_task->GetKeyForTbeKernel(), EXT_ATTR_ATOMIC_TBE_KERNEL);
+  EXPECT_EQ(atomic_task->GetKeyForTvmMagic(), ATOMIC_ATTR_TVM_MAGIC);
+  EXPECT_EQ(atomic_task->GetKeyForTvmMetaData(), ATOMIC_ATTR_TVM_METADATA);
+  EXPECT_EQ(atomic_task->GetKeyForKernelName(op_desc), "Sum_atomic_kernelname");
+}