!1735 shared memory optimize for wide&deep

From: @isaacxr
Reviewed-by: @xchu42,@wqtshg
Signed-off-by: @wqtshg

ge/graph/load/model_manager/davinci_model.cc

@@ -3727,6 +3727,8 @@ Status DavinciModel::InitTbeHandle(const OpDescPtr &op_desc) {
         binary.magic = RT_DEV_BINARY_MAGIC_ELF;
       } else if (json_string == "RT_DEV_BINARY_MAGIC_ELF_AIVEC") {
         binary.magic = RT_DEV_BINARY_MAGIC_ELF_AIVEC;
+      } else if (json_string == "RT_DEV_BINARY_MAGIC_ELF_AICUBE") {
+        binary.magic = RT_DEV_BINARY_MAGIC_ELF_AICUBE;
       } else {
         REPORT_INNER_ERROR("E19999", "Attr:%s value:%s in op:%s(%s), model_id:%u, check invalid",
                            TVM_ATTR_NAME_MAGIC.c_str(), json_string.c_str(),

ge/graph/manager/host_mem_manager.cc

@@ -107,16 +107,15 @@ Status HostMemManager::MallocSharedMemory(SharedMemInfo &mem_info) {
   return SUCCESS;
 }
 
-Status HostMemManager::QueryVarMemInfo(const string &op_name, uint64_t &base_addr, uint64_t &data_size) {
+bool HostMemManager::QueryVarMemInfo(const string &op_name, SharedMemInfo &mem_info) {
   std::lock_guard<std::recursive_mutex> lock(mutex_);
-  if (var_memory_base_map_.find(op_name) == var_memory_base_map_.end()) {
-    REPORT_INNER_ERROR("E19999", "MemInfo.op_name:%s can't find in var_memory_base_map_", op_name.c_str());
-    GELOGE(INTERNAL_ERROR, "[Check][Param] Find host base base_addr failed, node name:%s!", op_name.c_str());
-    return INTERNAL_ERROR;
+  auto it = var_memory_base_map_.find(op_name);
+  if (it == var_memory_base_map_.end()) {
+    GELOGW("Host memory for node [%s] not found.", op_name.c_str());
+    return false;
   }
-  base_addr = static_cast<uint64_t>(reinterpret_cast<uintptr_t>(var_memory_base_map_[op_name].device_address));
-  data_size = var_memory_base_map_[op_name].mem_size;
-  return SUCCESS;
+  mem_info = it->second;
+  return true;
 }
 
 string HostMemManager::OpNameToShmName(const string &op_name) {

ge/graph/manager/host_mem_manager.h

@@ -66,7 +66,7 @@ class HostMemManager {
   Status Initialize();
   void Finalize() noexcept;
   Status MallocSharedMemory(SharedMemInfo &mem_nfo);
-  Status QueryVarMemInfo(const string &op_name, uint64_t &base_addr, uint64_t &data_size);
+  bool QueryVarMemInfo(const string &op_name, SharedMemInfo &mem_info);
 
  private:
   static string OpNameToShmName(const string &op_name);

ge/graph/manager/memory_api.cc

@@ -110,7 +110,14 @@ Status MallocSharedMemory(const TensorInfo &tensor_info, uint64_t &dev_addr, uin
 }
 
 Status GetVarBaseAddrAndSize(const string &var_name, uint64_t &base_addr, uint64_t &var_size) {
-  GELOGD("GetVarBaseAddrAndSize in");
-  return HostMemManager::Instance().QueryVarMemInfo(var_name, base_addr, var_size);
+  GELOGD("GetVarBaseAddrAndSize in, var name:[%s]", var_name.c_str());
+  SharedMemInfo mem_info;
+  if (!HostMemManager::Instance().QueryVarMemInfo(var_name, mem_info)) {
+    GELOGE(FAILED, "Get addr and size failed, name:[%s]", var_name.c_str());
+    return FAILED;
+  }
+  base_addr = static_cast<uint64_t>(reinterpret_cast<uintptr_t>(mem_info.host_aligned_ptr->Get()));
+  var_size = mem_info.mem_size;
+  return SUCCESS;
 }
 }  // namespace ge

ge/hybrid/model/hybrid_model_builder.cc

@@ -1074,21 +1074,25 @@ Status HybridModelBuilder::InitVariableTensors() {
       GELOGE(INTERNAL_ERROR, "[Calculate][TensorMemSize] failed, node name:%s", it.first.c_str());
       return INTERNAL_ERROR;
     }
-    SharedMemInfo mem_info(it.first, tensor_size);
-    if (HostMemManager::Instance().MallocSharedMemory(mem_info) != SUCCESS) {
-      GELOGE(GE_GRAPH_MALLOC_FAILED, "[Malloc][SharedMemory] failed, Host variable [%s].", it.first.c_str());
-      return GE_GRAPH_MALLOC_FAILED;
-    }
-    if (MemManager::Instance().HostMemInstance(RT_MEMORY_HBM).Malloc(mem_info.host_aligned_ptr,
-                                                                     tensor_size) == nullptr) {
-      GELOGE(MEMALLOC_FAILED, "[Malloc][HostMem] for an existed GeTensor failed, Host variable [%s].",
-             it.first.c_str());
+
+    // Host variable will be assigned to allocated shared memory first.
+    SharedMemInfo mem_info;
+    void *mem_addr = nullptr;
+    if (HostMemManager::Instance().QueryVarMemInfo(it.first, mem_info)) {
+      mem_addr = const_cast<void *>(MemManager::Instance().HostMemInstance(RT_MEMORY_HBM)
+                                      .Malloc(mem_info.host_aligned_ptr, tensor_size));
+    } else {
+      mem_addr = MemManager::Instance().HostMemInstance(RT_MEMORY_HBM).Malloc(tensor_size);
+    }
+
+    if (mem_addr == nullptr) {
+      REPORT_INNER_ERROR("E19999", "[Malloc][HostMem] for variable [%s] failed.", it.first.c_str());
+      GELOGE(MEMALLOC_FAILED, "[Malloc][HostMem] for variable [%s] failed.", it.first.c_str());
       return MEMALLOC_FAILED;
     }
     GELOGD("Host variable [%s] malloc success, size=%ld.", it.first.c_str(), tensor_size);
 
-    std::unique_ptr<TensorValue> tensor(new (std::nothrow) TensorValue(mem_info.host_aligned_ptr->MutableGet(),
-                                                                       tensor_size));
+    std::unique_ptr<TensorValue> tensor(new (std::nothrow) TensorValue(mem_addr, tensor_size));
     GE_CHECK_NOTNULL(tensor);
     hybrid_model_.variable_tensors_.emplace(it.first, std::move(tensor));
   }

tests/ut/ge/CMakeLists.txt

@@ -157,6 +157,7 @@ set(COMMON_SRC_FILES
     "${GE_CODE_DIR}/ge/graph/load/model_manager/model_manager.cc"
     "${GE_CODE_DIR}/ge/common/profiling/profiling_manager.cc"
     "${GE_CODE_DIR}/ge/graph/manager/host_mem_manager.cc"
+	"${GE_CODE_DIR}/ge/graph/manager/memory_api.cc"
     "${GE_CODE_DIR}/ge/session/inner_session.cc"
     "${GE_CODE_DIR}/ge/graph/manager/util/rt_context_util.cc"
     "${GE_CODE_DIR}/ge/graph/execute/graph_execute.cc"
@@ -799,6 +800,7 @@ set(MULTI_PARTS_TEST_FILES
     "graph/manager/hcom_util_unittest.cc"
     "graph/manager/graph_caching_allocator_unittest.cc"
     "graph/manager/host_mem_allocator_unittest.cc"
+	"graph/manager/memory_api_unittest.cc"
     "graph/manager/session_scope_mem_allocator_unittest.cc"
     "graph/manager/run_graph_unittest.cc"
     "graph/partition/dynamic_shape_partition_unittest.cc"

tests/ut/ge/graph/manager/memory_api_unittest.cc

@@ -0,0 +1,68 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <gtest/gtest.h>
+#include <memory>
+
+
+
+#define protected public
+#define private public
+#include "graph/manager/host_mem_manager.h"
+#include "inc/framework/memory/memory_api.h"
+#undef protected
+#undef private
+#include "metadef/inc/graph/aligned_ptr.h"
+
+using namespace std;
+using namespace testing;
+using namespace ge;
+
+class UtestMemoryApiTest : public testing::Test {
+ protected:
+  void SetUp() {}
+
+  void TearDown() {}
+};
+
+TEST_F(UtestMemoryApiTest, query_mem_info_success) {
+string var_name = "host_params";
+SharedMemInfo info;
+uint8_t tmp(0);
+info.device_address = &tmp;
+
+std::shared_ptr<AlignedPtr> aligned_ptr = std::make_shared<AlignedPtr>(100, 16);
+
+info.host_aligned_ptr = aligned_ptr;
+info.fd=0;
+info.mem_size = 100;
+info.op_name = var_name;
+HostMemManager::Instance().var_memory_base_map_[var_name] = info;
+uint64_t base_addr;
+uint64_t var_size;
+Status ret = GetVarBaseAddrAndSize(var_name, base_addr, var_size);
+EXPECT_EQ(ret, SUCCESS);
+EXPECT_EQ(var_size, 100);
+HostMemManager::Instance().var_memory_base_map_.clear();
+}
+
+TEST_F(UtestMemoryApiTest, query_mem_info_failed) {
+string var_name = "host_params";
+uint64_t base_addr;
+uint64_t var_size;
+Status ret = GetVarBaseAddrAndSize(var_name, base_addr, var_size);
+EXPECT_NE(ret, SUCCESS);
+}

tests/ut/ge/hybrid/model/hybrid_model_builder_unittest.cc

@@ -23,6 +23,7 @@
 #define protected public
 #include "hybrid/model/hybrid_model_builder.h"
 #include "hybrid/node_executor/node_executor.h"
+#include "graph/manager/host_mem_manager.h"
 
 #include "graph/utils/tensor_utils.h"
 #include "graph/utils/graph_utils.h"
@@ -263,4 +264,59 @@ TEST_F(UtestHybridModelBuilder, init_constant_op_host_) {
   EXPECT_EQ(hybrid_model_builder.InitConstantOps(), SUCCESS);
   EXPECT_EQ(hybrid_model_builder.hybrid_model_.variable_tensors_.size(), 2);
 }
+
+TEST_F(UtestHybridModelBuilder, init_host_var_with_host_mem) {
+ComputeGraphPtr graph = std::make_shared<ComputeGraph>("test");
+GeRootModelPtr ge_root_model = make_shared<GeRootModel>(graph);
+HybridModel hybrid_model(ge_root_model);
+HybridModelBuilder hybrid_model_builder(hybrid_model);
+
+OpDescPtr op_desc = std::make_shared<OpDesc>("host_params", VARIABLE);
+GeTensorDesc tensor_desc(GeShape(),FORMAT_NHWC,DT_FLOAT);
+TensorUtils::SetSize(tensor_desc, 512);
+op_desc->AddOutputDesc(tensor_desc);
+auto host_var = graph->AddNode(op_desc);
+
+hybrid_model.host_variable_nodes_.emplace("host_params", host_var);
+std::map<std::string, string> options;
+options["ge.exec.placement"] = "HOST";
+GetThreadLocalContext().SetGraphOption(options);
+
+EXPECT_EQ(hybrid_model_builder.InitVariableTensors(), SUCCESS);
+EXPECT_EQ(hybrid_model_builder.hybrid_model_.variable_tensors_.size(), 1);
+}
+
+TEST_F(UtestHybridModelBuilder, init_host_var_with_host_shared_mem) {
+ComputeGraphPtr graph = std::make_shared<ComputeGraph>("test");
+GeRootModelPtr ge_root_model = make_shared<GeRootModel>(graph);
+HybridModel hybrid_model(ge_root_model);
+HybridModelBuilder hybrid_model_builder(hybrid_model);
+
+OpDescPtr op_desc = std::make_shared<OpDesc>("host_params", VARIABLE);
+GeTensorDesc tensor_desc(GeShape(),FORMAT_NHWC,DT_FLOAT);
+TensorUtils::SetSize(tensor_desc, 512);
+op_desc->AddOutputDesc(tensor_desc);
+auto host_var = graph->AddNode(op_desc);
+
+hybrid_model.host_variable_nodes_.emplace("host_params", host_var);
+std::map<std::string, string> options;
+options["ge.exec.placement"] = "HOST";
+GetThreadLocalContext().SetGraphOption(options);
+
+SharedMemInfo info;
+uint8_t tmp(0);
+info.device_address = &tmp;
+std::shared_ptr<AlignedPtr> aligned_ptr = std::make_shared<AlignedPtr>(512, 16);
+info.host_aligned_ptr = aligned_ptr;
+info.fd=0;
+info.mem_size = 100;
+info.op_name = "host_params";
+HostMemManager::Instance().var_memory_base_map_["host_params"] = info;
+
+
+
+EXPECT_EQ(hybrid_model_builder.InitVariableTensors(), SUCCESS);
+EXPECT_EQ(hybrid_model_builder.hybrid_model_.variable_tensors_.size(), 1);
+HostMemManager::Instance().var_memory_base_map_.clear();
+}
 } // namespace ge