!9684 remove pybind calling in cxx library

From: @zhoufeng54
Reviewed-by: @xu-yfei,@chujinjin
Signed-off-by: @chujinjin

cmake/package.cmake

@@ -37,6 +37,7 @@ set(INSTALL_LIB_DIR ${CMAKE_INSTALL_LIBDIR} CACHE PATH "Installation directory f
 set(INSTALL_PY_DIR ".")
 set(INSTALL_BASE_DIR ".")
 set(INSTALL_BIN_DIR "bin")
+set(INSTALL_CFG_DIR "config")
 
 if (CMAKE_SYSTEM_NAME MATCHES "Windows")
     set(INSTALL_LIB_DIR ".")
@@ -308,6 +309,13 @@ install(
     COMPONENT mindspore
 )
 
+## config files
+install(
+    FILES ${CMAKE_SOURCE_DIR}/config/op_info.config
+    DESTINATION ${INSTALL_CFG_DIR}
+    COMPONENT mindspore
+)
+
 if (ENABLE_SERVING)
     install(
         TARGETS ms_serving

include/api/types.h

@@ -106,17 +106,15 @@ extern MS_API const char *kDeviceTypeAscend310;
 extern MS_API const char *kDeviceTypeAscend910;
 
 constexpr auto kModelOptionDumpCfgPath = "mindspore.option.dump_config_file_path";
-constexpr auto kModelOptionDvppCfgPath = "mindspore.option.dvpp_config_file_path";
 constexpr auto kModelOptionInsertOpCfgPath = "mindspore.option.insert_op_config_file_path";  // aipp config file
 constexpr auto kModelOptionInputFormat = "mindspore.option.input_format";                    // nchw or nhwc
 // Mandatory while dynamic batch: e.g. "input_op_name1: n1,c2,h3,w4;input_op_name2: n4,c3,h2,w1"
 constexpr auto kModelOptionInputShape = "mindspore.option.input_shape";
-constexpr auto kModelOptionDynamicBatchSize = "mindspore.option.dynamic_batch_size";
-constexpr auto kModelOptionDynamicImageSize = "mindspore.option.dynamic_image_size";
-constexpr auto kModelOptionDynamicDims = "mindspore.option.dynamic_dims";
-constexpr auto kModelOptionSerialInput = "mindspore.option.serial_inputs_name";  // separated by ';'
-constexpr auto kModelOptionOutputNode = "mindspore.option.output_node";          // e.g. "node_name1:0;node_name2:1"
 constexpr auto kModelOptionOutputType = "mindspore.option.output_type";  // "FP32", "UINT8" or "FP16", default as "FP32"
+constexpr auto kModelOptionPrecisionMode = "mindspore.option.precision_mode";
+// "force_fp16", "allow_fp32_to_fp16", "must_keep_origin_dtype" or "allow_mix_precision", default as "force_fp16"
+constexpr auto kModelOptionOpSelectImplMode = "mindspore.option.op_select_impl_mode";
+// "high_precision" or "high_performance", default as "high_performance"
 }  // namespace api
 }  // namespace mindspore
 #endif  // MINDSPORE_INCLUDE_API_TYPES_H

mindspore/ccsrc/cxx_api/CMakeLists.txt

@@ -10,6 +10,7 @@ if (ENABLE_ACL)
     include_directories(${CMAKE_SOURCE_DIR}/graphengine/src/ge)
     include_directories(${CMAKE_BINARY_DIR}/proto/ge)
     file(GLOB_RECURSE API_ACL_SRC ${CMAKE_CURRENT_SOURCE_DIR}
+            "python_utils.cc"
             "model/acl/*.cc"
             "model/model_converter_utils/*.cc"
             "graph/acl/*.cc"
@@ -17,14 +18,13 @@ if (ENABLE_ACL)
 
 endif ()
 if (ENABLE_D)
-    file(GLOB_RECURSE API_MS_INFER_SRC ${CMAKE_CURRENT_SOURCE_DIR} "model/ms/*.cc" "graph/ms/*.cc")
+    file(GLOB_RECURSE API_MS_INFER_SRC ${CMAKE_CURRENT_SOURCE_DIR} "python_utils.cc" "model/ms/*.cc" "graph/ms/*.cc")
 endif ()
 
 set(MSLIB_SRC ${CMAKE_CURRENT_SOURCE_DIR}/types.cc
               ${CMAKE_CURRENT_SOURCE_DIR}/context.cc
               ${CMAKE_CURRENT_SOURCE_DIR}/cell.cc
               ${CMAKE_CURRENT_SOURCE_DIR}/serialization.cc
-              ${CMAKE_CURRENT_SOURCE_DIR}/python_utils.cc
               ${CMAKE_CURRENT_SOURCE_DIR}/graph/graph.cc
               ${CMAKE_CURRENT_SOURCE_DIR}/graph/graph_data.cc
               ${CMAKE_CURRENT_SOURCE_DIR}/model/model.cc

mindspore/ccsrc/cxx_api/graph/acl/acl_graph_impl.cc

@@ -219,7 +219,6 @@ Status AclGraphImpl::Load() {
 
 Status AclGraphImpl::ConvertToOM() {
   MS_LOG(INFO) << "Start convert to om model.";
-  RegAllOpFromPython();
   if (graph_ == nullptr) {
     MS_LOG(ERROR) << "Invalid graph_ is null.";
     return FAILED;

mindspore/ccsrc/cxx_api/graph/acl/model_process.cc

@@ -15,6 +15,7 @@
  */
 
 #include "cxx_api/graph/acl/model_process.h"
+#include <sys/time.h>
 #include <algorithm>
 #include <map>
 #include "utils/utils.h"
@@ -345,7 +346,18 @@ Status ModelProcess::PredictFromHost(const std::vector<Buffer> &inputs, std::vec
     DestroyInputsDataset();
     return ret;  // forward status error
   }
+
+  struct timeval start_time;
+  struct timeval end_time;
+  (void)gettimeofday(&start_time, nullptr);
   acl_ret = aclmdlExecute(model_id_, inputs_, outputs_);
+  (void)gettimeofday(&end_time, nullptr);
+  constexpr uint64_t kUSecondInSecond = 1000000;
+  uint64_t cost =
+    (kUSecondInSecond * static_cast<uint64_t>(end_time.tv_sec) + static_cast<uint64_t>(end_time.tv_usec)) -
+    (kUSecondInSecond * static_cast<uint64_t>(start_time.tv_sec) + static_cast<uint64_t>(start_time.tv_usec));
+  MS_LOG(INFO) << "Model execute in " << cost << " us";
+
   DestroyInputsDataset();
   if (acl_ret != ACL_ERROR_NONE) {
     MS_LOG(ERROR) << "Execute Model Failed";

mindspore/ccsrc/cxx_api/graph/ms/ms_graph_impl.cc

@@ -72,12 +72,16 @@ MsGraphImpl::MsGraphImpl()
       outputs_(),
       input_names_(),
       output_names_(),
+      init_flag_(false),
       load_flag_(false) {}
 
 MsGraphImpl::~MsGraphImpl() { (void)FinalizeEnv(); }
 
 Status MsGraphImpl::InitEnv() {
-  RegAllOpFromPython();
+  if (init_flag_) {
+    return SUCCESS;
+  }
+  RegAllOp();
   auto ms_context = MsContext::GetInstance();
   if (ms_context == nullptr) {
     MS_LOG(ERROR) << "Get Context failed!";
@@ -100,12 +104,16 @@ Status MsGraphImpl::InitEnv() {
   }
   session_impl_->Init(device_id_);
 
+  init_flag_ = true;
   return SUCCESS;
 }
 
 Status MsGraphImpl::FinalizeEnv() {
+  if (!init_flag_) {
+    return SUCCESS;
+  }
+
   MS_LOG_INFO << "Start finalize env";
-  pybind11::gil_scoped_acquire acquire;
   session::ExecutorManager::Instance().Clear();
   device::KernelRuntimeManager::Instance().ClearRuntimeResource();
   auto ms_context = MsContext::GetInstance();
@@ -117,6 +125,8 @@ Status MsGraphImpl::FinalizeEnv() {
     MS_LOG(ERROR) << "CloseTsd failed!";
     return FAILED;
   }
+
+  init_flag_ = false;
   MS_LOG(INFO) << "End finalize env";
   return SUCCESS;
 }
@@ -125,7 +135,6 @@ Status MsGraphImpl::CompileGraph(const std::shared_ptr<FuncGraph> &funcGraphPtr)
   MS_ASSERT(session_impl_ != nullptr);
   try {
     graph_id_ = session_impl_->CompileGraph(NOT_NULL(funcGraphPtr));
-    pybind11::gil_scoped_release gil_release;
     return SUCCESS;
   } catch (std::exception &e) {
     MS_LOG(ERROR) << "CompileGraph failed: " << e.what();

mindspore/ccsrc/cxx_api/graph/ms/ms_graph_impl.h

@@ -59,6 +59,7 @@ class MsGraphImpl : public GraphCell::GraphImpl {
   std::vector<tensor::TensorPtr> outputs_;
   std::vector<std::string> input_names_;
   std::vector<std::string> output_names_;
+  bool init_flag_;
   bool load_flag_;
 };
 }  // namespace mindspore::api

mindspore/ccsrc/cxx_api/model/acl/acl_model.cc

@@ -25,7 +25,6 @@ API_FACTORY_REG(ModelImpl, Ascend310, AclModel);
 Status AclModel::Build(const std::map<std::string, std::string> &options_map) {
   MS_LOG(INFO) << "Start build model.";
   MS_EXCEPTION_IF_NULL(graph_);
-  RegAllOpFromPython();
   std::unique_ptr<AclModelOptions> options = std::make_unique<AclModelOptions>(options_map);
   std::string options_str = GenerateOptionsStr(options_map);
   MS_EXCEPTION_IF_NULL(options);

mindspore/ccsrc/cxx_api/model/acl/acl_model_options.cc

@@ -15,6 +15,7 @@
  */
 #include "cxx_api/model/acl/acl_model_options.h"
 #include <memory>
+#include "utils/log_adapter.h"
 #include "external/ge/ge_api_types.h"
 
 namespace mindspore::api {
@@ -27,40 +28,45 @@ static std::string ParseOption(const std::map<std::string, std::string> &options
 }
 
 AclModelOptions::AclModelOptions(const std::map<std::string, std::string> &options) {
-  dump_cfg_path = ParseOption(options, kModelOptionDumpCfgPath);
-  dvpp_cfg_path = ParseOption(options, kModelOptionDvppCfgPath);
-  output_node = ParseOption(options, kModelOptionOutputNode);
   // to acl
   insert_op_cfg_path = ParseOption(options, kModelOptionInsertOpCfgPath);
   input_format = ParseOption(options, kModelOptionInputFormat);
   input_shape = ParseOption(options, kModelOptionInputShape);
-  dynamic_batch_size = ParseOption(options, kModelOptionInputShape);
-  dynamic_image_size = ParseOption(options, kModelOptionInputShape);
-  dynamic_dims = ParseOption(options, kModelOptionInputShape);
-  serial_nodes_name = ParseOption(options, kModelOptionSerialInput);
   output_type = ParseOption(options, kModelOptionOutputType);
+  precision_mode = ParseOption(options, kModelOptionPrecisionMode);
+  op_select_impl_mode = ParseOption(options, kModelOptionOpSelectImplMode);
 }
 
-std::map<std::string, std::string> AclModelOptions::GenAclOptions() const {
-  const std::map<std::string const *, std::string> acl_options_map = {
-    {&insert_op_cfg_path, ge::ir_option::INSERT_OP_FILE},
-    {&input_format, ge::ir_option::INPUT_FORMAT},
-    {&input_shape, ge::ir_option::INPUT_SHAPE},
-    {&dynamic_batch_size, ge::ir_option::DYNAMIC_BATCH_SIZE},
-    {&dynamic_image_size, ge::ir_option::DYNAMIC_IMAGE_SIZE},
-    {&dynamic_dims, ge::ir_option::DYNAMIC_DIMS},
-    {&serial_nodes_name, ge::ir_option::INPUT_FP16_NODES},
-    {&output_type, ge::ir_option::OUTPUT_TYPE},
+std::tuple<std::map<std::string, std::string>, std::map<std::string, std::string>> AclModelOptions::GenAclOptions()
+  const {
+  const std::map<std::string const *, std::string> init_options_map = {
+    {&op_select_impl_mode, ge::ir_option::OP_SELECT_IMPL_MODE},
+    {&soc_version, ge::ir_option::SOC_VERSION},
   };
 
-  std::map<std::string, std::string> acl_options;
-  for (auto [ms_option, acl_option_key] : acl_options_map) {
+  const std::map<std::string const *, std::string> build_options_map = {
+    {&insert_op_cfg_path, ge::ir_option::INSERT_OP_FILE}, {&input_format, ge::ir_option::INPUT_FORMAT},
+    {&input_shape, ge::ir_option::INPUT_SHAPE},           {&output_type, ge::ir_option::OUTPUT_TYPE},
+    {&precision_mode, ge::ir_option::PRECISION_MODE},
+  };
+
+  std::map<std::string, std::string> init_options;
+  std::map<std::string, std::string> build_options;
+  for (auto [ms_option, acl_option_key] : init_options_map) {
     if (ms_option == nullptr || ms_option->empty()) {
       continue;
     }
-    acl_options.emplace(acl_option_key, *ms_option);
+    MS_LOG(INFO) << "Option " << acl_option_key << " : " << *ms_option;
+    init_options.emplace(acl_option_key, *ms_option);
   }
-  return acl_options;
-}
 
+  for (auto [ms_option, acl_option_key] : build_options_map) {
+    if (ms_option == nullptr || ms_option->empty()) {
+      continue;
+    }
+    MS_LOG(INFO) << "Option " << acl_option_key << " : " << *ms_option;
+    build_options.emplace(acl_option_key, *ms_option);
+  }
+  return {init_options, build_options};
+}
 }  // namespace mindspore::api

mindspore/ccsrc/cxx_api/model/acl/acl_model_options.h

@@ -19,28 +19,27 @@
 #include <vector>
 #include <string>
 #include <map>
+#include <tuple>
 #include "include/api/types.h"
 #include "include/api/status.h"
 
 namespace mindspore::api {
 struct AclModelOptions {
-  std::string dump_cfg_path;
-  std::string dvpp_cfg_path;
   std::string output_node;  // todo: at convert.cc::BuildGraph(), no atc options
   // build options
   std::string insert_op_cfg_path;
   std::string input_format;
   std::string input_shape;
-  std::string dynamic_batch_size;
-  std::string dynamic_image_size;
-  std::string dynamic_dims;
-  std::string serial_nodes_name;
   std::string output_type;
+  std::string precision_mode;
+  std::string op_select_impl_mode;
+  std::string soc_version = "Ascend310";
 
   explicit AclModelOptions(const std::map<std::string, std::string> &options);
   ~AclModelOptions() = default;
 
-  std::map<std::string, std::string> GenAclOptions() const;
+  // return tuple<init_options, build_options>
+  std::tuple<std::map<std::string, std::string>, std::map<std::string, std::string>> GenAclOptions() const;
 };
 }  // namespace mindspore::api
 

mindspore/ccsrc/cxx_api/model/acl/model_converter.cc

@@ -131,17 +131,16 @@ transform::DfGraphPtr ModelConverter::ConvertFuncGraphToAIR(const FuncGraphPtr &
 }
 
 Buffer ModelConverter::BuildAirModel(const transform::DfGraphPtr &graph,
-                                     const std::map<std::string, std::string> &acl_options) {
+                                     const std::map<std::string, std::string> &init_options,
+                                     const std::map<std::string, std::string> &build_options) {
   ge::ModelBufferData model;
-  auto ge_options = acl_options;
-  ge_options.emplace(ge::ir_option::SOC_VERSION, "Ascend310");
-  auto ret = ge::aclgrphBuildInitialize(ge_options);
+  auto ret = ge::aclgrphBuildInitialize(init_options);
   if (ret != ge::SUCCESS) {
     MS_LOG(ERROR) << "Call aclgrphBuildInitialize fail.";
     return Buffer();
   }
 
-  ret = ge::aclgrphBuildModel(*graph, acl_options, model);
+  ret = ge::aclgrphBuildModel(*graph, build_options, model);
   if (ret != ge::SUCCESS) {
     MS_LOG(ERROR) << "Call aclgrphBuildModel fail.";
     return Buffer();
@@ -290,7 +289,6 @@ Buffer ModelConverter::LoadAscendIR(const Buffer &model_data) {
 }
 
 Buffer ModelConverter::LoadMindIRInner(const FuncGraphPtr &func_graph) {
-  RegAllOpFromPython();
   if (func_graph == nullptr) {
     MS_LOG(ERROR) << "Convert MindIR to FuncGraph failed.";
     return Buffer();
@@ -302,17 +300,17 @@ Buffer ModelConverter::LoadMindIRInner(const FuncGraphPtr &func_graph) {
     return Buffer();
   }
 
-  std::map<std::string, std::string> acl_options;
+  std::map<std::string, std::string> init_options;
+  std::map<std::string, std::string> build_options;
   if (options_ != nullptr) {
-    acl_options = options_->GenAclOptions();
+    std::tie(init_options, build_options) = options_->GenAclOptions();
   }
 
-  auto om_data = BuildAirModel(df_graph, acl_options);
+  auto om_data = BuildAirModel(df_graph, init_options, build_options);
   return om_data;
 }
 
 Buffer ModelConverter::LoadAscendIRInner(const Buffer &model_data) {
-  RegAllOpFromPython();
   ge::Model load_model = ge::Model("loadmodel", "version2");
   ge::Status ret =
     ge::Model::Load(reinterpret_cast<const uint8_t *>(model_data.Data()), model_data.DataSize(), load_model);
@@ -327,12 +325,13 @@ Buffer ModelConverter::LoadAscendIRInner(const Buffer &model_data) {
     return Buffer();
   }
 
-  std::map<std::string, std::string> acl_options;
+  std::map<std::string, std::string> init_options;
+  std::map<std::string, std::string> build_options;
   if (options_ != nullptr) {
-    acl_options = options_->GenAclOptions();
+    std::tie(init_options, build_options) = options_->GenAclOptions();
   }
 
-  auto om_data = BuildAirModel(df_graph, acl_options);
+  auto om_data = BuildAirModel(df_graph, init_options, build_options);
   return om_data;
 }
 }  // namespace mindspore::api

mindspore/ccsrc/cxx_api/model/acl/model_converter.h

@@ -39,7 +39,8 @@ class ModelConverter {
 
  private:
   transform::DfGraphPtr ConvertFuncGraphToAIR(const FuncGraphPtr &anf_graph);
-  Buffer BuildAirModel(const transform::DfGraphPtr &graph, const std::map<std::string, std::string> &acl_options);
+  Buffer BuildAirModel(const transform::DfGraphPtr &graph, const std::map<std::string, std::string> &init_options,
+                       const std::map<std::string, std::string> &build_options);
   AclModelOptions *options_;
 
   Buffer LoadMindIRInner(const FuncGraphPtr &func_graph);

mindspore/ccsrc/cxx_api/python_utils.cc

@@ -14,25 +14,23 @@
  * limitations under the License.
  */
 #include "cxx_api/python_utils.h"
+#include <dlfcn.h>
 #include <mutex>
 #include <vector>
 #include <memory>
+#include <string>
+#include <fstream>
 #include "mindspore/core/utils/ms_context.h"
 #include "pybind11/pybind11.h"
 #include "backend/kernel_compiler/oplib/oplib.h"
 
 namespace py = pybind11;
 
-namespace mindspore::api {
-void RegAllOpFromPython() {
-  static std::mutex init_mutex;
-  static bool Initialized = false;
+static std::mutex init_mutex;
+static bool Initialized = false;
 
-  std::lock_guard<std::mutex> lock(init_mutex);
-  if (Initialized) {
-    return;
-  }
-  Initialized = true;
+namespace mindspore::api {
+static void RegAllOpFromPython() {
   MsContext::GetInstance()->set_param<int>(MS_CTX_EXECUTION_MODE, kGraphMode);
   Py_Initialize();
   auto c_expression = PyImport_ImportModule("mindspore._c_expression");
@@ -61,5 +59,64 @@ void RegAllOpFromPython() {
   Py_DECREF(c_expression);
 }
 
+static bool RegAllOpFromFile() {
+  Dl_info info;
+  int dl_ret = dladdr(reinterpret_cast<void *>(RegAllOpFromFile), &info);
+  if (dl_ret == 0) {
+    MS_LOG(INFO) << "Get dladdr failed, skip.";
+    return false;
+  }
+  std::string dir(info.dli_fname);
+  MS_LOG(INFO) << "Get library path is " << dir;
+
+  auto split_pos = dir.find_last_of('/');
+  if (dir.empty() || split_pos == std::string::npos) {
+    MS_LOG(INFO) << "Missing op config file, skip.";
+    return false;
+  }
+
+  dir = dir.substr(0, split_pos) + "/../config/op_info.config";
+  if (dir.size() >= PATH_MAX) {
+    MS_LOG(ERROR) << "Op info path is invalid: " << dir;
+    return false;
+  }
+
+  char real_path_mem[PATH_MAX] = {0};
+  if (realpath(common::SafeCStr(dir), real_path_mem) == nullptr) {
+    MS_LOG(ERROR) << "Op info path is invalid: " << dir;
+    return false;
+  }
+  std::string real_path(real_path_mem);
+
+  MS_LOG(INFO) << "Start to read op info from local file " << real_path;
+  std::ifstream file(real_path);
+  if (!file.is_open()) {
+    MS_LOG(ERROR) << "Find op info file failed.";
+    return false;
+  }
+  std::string line;
+  while (getline(file, line)) {
+    if (!line.empty()) {
+      (void)kernel::OpLib::RegOp(line, "");
+    }
+  }
+  MS_LOG(INFO) << "End";
+  return true;
+}
+
+void RegAllOp() {
+  std::lock_guard<std::mutex> lock(init_mutex);
+  if (Initialized) {
+    return;
+  }
+  bool ret = RegAllOpFromFile();
+  if (!ret) {
+    MS_LOG(INFO) << "Reg all op from file failed, start to reg from python.";
+    RegAllOpFromPython();
+  }
+
+  Initialized = true;
+}
+
 bool PythonIsInited() { return Py_IsInitialized() != 0; }
 }  // namespace mindspore::api

mindspore/ccsrc/cxx_api/python_utils.h

@@ -17,10 +17,8 @@
 #ifndef MINDSPORE_CCSRC_CXXAPI_PYTHON_UTILS_H
 #define MINDSPORE_CCSRC_CXXAPI_PYTHON_UTILS_H
 
-#include "pybind11/pybind11.h"
-
 namespace mindspore::api {
-void RegAllOpFromPython();
+void RegAllOp();
 bool PythonIsInited();
 }  // namespace mindspore::api
 

setup.py

@@ -133,6 +133,7 @@ package_data = {
         'lib/*.a',
         '.commit_id',
         'ms_serving',
+        'config/*'
         'include/*',
         'include/*/*',
         'include/*/*/*',

tests/cxx_st/CMakeLists.txt

@@ -8,4 +8,4 @@ include_directories(${CUDA_INCLUDE_DIRS})
 
 file(GLOB_RECURSE CXX_ST_SRC RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} *.cc)
 add_executable(st_tests ${CXX_ST_SRC})
-target_link_libraries(st_tests PRIVATE mindspore_shared_lib mindspore::gtest)
+target_link_libraries(st_tests PRIVATE mindspore_shared_lib mindspore::gtest)