!8965 Synchronize latest Ascend software suite 09 Dec 2020, with refactored GraphEngine

From: @nicholas_yhr
Reviewed-by: 
Signed-off-by:

CMakeLists.txt

@@ -71,7 +71,7 @@ set(MS_CCSRC_PATH ${CMAKE_SOURCE_DIR}/mindspore/ccsrc)
 set(MS_CCSRC_BUILD_PATH ${BUILD_PATH}/mindspore/mindspore/ccsrc)
 
 if (ENABLE_D OR ENABLE_ACL OR ENABLE_TESTCASES)
-    include(${CMAKE_SOURCE_DIR}/cmake/dependency_graphengine.cmake)
+    include(${CMAKE_SOURCE_DIR}/cmake/dependency_graphenginev2.cmake)
 endif ()
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fvisibility=hidden")

build.sh

@@ -398,6 +398,9 @@ parse_device
 echo "---------------- MindSpore: build start ----------------"
 mkdir -pv "${BUILD_PATH}/package/mindspore/lib"
 git submodule update --init graphengine
+cd "${BASEPATH}/graphengine"
+git submodule update --init metadef
+cd "${BASEPATH}"
 if [[ "X$ENABLE_AKG" = "Xon" ]] && [[ "X$ENABLE_D" = "Xon" || "X$ENABLE_GPU" = "Xon" ]]; then
     git submodule update --init --recursive akg
 fi

cmake/dependency_graphenginev2.cmake

@@ -0,0 +1,39 @@
+message(STATUS "Compiling GraphEngine")
+set(GE_SOURCE_DIR ${CMAKE_SOURCE_DIR}/graphengine)
+
+message(STATUS "[ME] build_path: ${BUILD_PATH}")
+
+function(find_submodule_lib module name path)
+    find_library(${module}_LIBRARY_DIR NAMES ${name} NAMES_PER_DIR PATHS ${path}
+            PATH_SUFFIXES lib
+            )
+    if ("${${module}_LIBRARY_DIR}" STREQUAL "${module}_LIBRARY_DIR-NOTFOUND")
+        message(FATAL_ERROR "${name} not found in any of following paths: ${path}")
+    endif()
+    add_library(${module} SHARED IMPORTED)
+    set_target_properties(${module} PROPERTIES
+            IMPORTED_LOCATION ${${module}_LIBRARY_DIR}
+            )
+endfunction()
+
+if (ENABLE_D OR ENABLE_ACL OR ENABLE_TESTCASES)
+    set(_ge_tmp_CMAKE_INSTALL_PREFIX ${CMAKE_INSTALL_PREFIX})
+    set(_ge_tmp_ENABLE_GITEE ${ENABLE_GITEE})
+    set(ENABLE_GITEE ON)
+    set(CMAKE_INSTALL_PREFIX ${BUILD_PATH}/graphengine)
+
+    if (ENABLE_TESTCASES)
+        # use slog, error manager, mmpa in non ascend mode, e.g. tests
+        set(GE_PREBUILD_PATH ${GE_SOURCE_DIR}/third_party/prebuild/${CMAKE_HOST_SYSTEM_PROCESSOR})
+        set(ENABLE_MS_TESTCASES TRUE)
+        find_submodule_lib(slog libslog.so ${GE_PREBUILD_PATH})
+        find_submodule_lib(error_manager liberror_manager.so ${GE_PREBUILD_PATH})
+        find_submodule_lib(static_mmpa libmmpa.a ${GE_PREBUILD_PATH})
+    endif()
+
+    add_subdirectory(${GE_SOURCE_DIR})
+    set(CMAKE_INSTALL_PREFIX ${_ge_tmp_CMAKE_INSTALL_PREFIX})
+    set(ENABLE_GITEE ${_ge_tmp_ENABLE_GITEE})
+else()
+    message(FATAL_ERROR "No compile option defined for GraphEngine, exiting")
+endif()

cmake/mind_expression.cmake

@@ -61,9 +61,10 @@ if (ENABLE_GE)
     link_directories(${CMAKE_SOURCE_DIR}/third_party/ge/lib)
 elseif(ENABLE_D OR ENABLE_ACL OR ENABLE_TESTCASES)
     include_directories(${CMAKE_SOURCE_DIR}/graphengine/inc)
-    include_directories(${CMAKE_SOURCE_DIR}/graphengine/inc/ops)
     include_directories(${CMAKE_SOURCE_DIR}/graphengine/inc/external)
-    include_directories(${CMAKE_SOURCE_DIR}/graphengine/inc/external/graph)
+    include_directories(${CMAKE_SOURCE_DIR}/graphengine/metadef/inc)
+    include_directories(${CMAKE_SOURCE_DIR}/graphengine/metadef/inc/external)
+    include_directories(${CMAKE_SOURCE_DIR}/graphengine/metadef/inc/external/graph)
 endif()
 
 if (ENABLE_GE OR ENABLE_D OR ENABLE_ACL OR ENABLE_TESTCASES)

cmake/package.cmake

@@ -211,7 +211,7 @@ if (NOT ENABLE_GE)
         set(ASCEND_DRIVER_PATH ${ASCEND_PATH}/driver/lib64/common)
 
         install(
-            FILES ${CMAKE_SOURCE_DIR}/build/graphengine/libc_sec.so
+            FILES ${CMAKE_SOURCE_DIR}/build/graphengine/c_sec/lib/libc_sec.so
             DESTINATION ${INSTALL_LIB_DIR}
             COMPONENT mindspore
         )
@@ -224,9 +224,9 @@ if (NOT ENABLE_GE)
             )
             install(
                 FILES
-                    ${CMAKE_BINARY_DIR}/graphengine/src/common/graph/libgraph.so
+                    ${CMAKE_BINARY_DIR}/graphengine/metadef/graph/libgraph.so
-                    ${CMAKE_BINARY_DIR}/graphengine/src/ge/common/libge_common.so
+                    ${CMAKE_BINARY_DIR}/graphengine/ge/common/libge_common.so
-                    ${CMAKE_BINARY_DIR}/graphengine/src/ge/ge_runtime/libge_runtime.so
+                    ${CMAKE_BINARY_DIR}/graphengine/ge/ge_runtime/libge_runtime.so
                 DESTINATION ${INSTALL_LIB_DIR}
                 COMPONENT mindspore
             )
@@ -234,8 +234,8 @@ if (NOT ENABLE_GE)
     elseif (ENABLE_TESTCASES)
         install(
             FILES
-                ${CMAKE_BINARY_DIR}/graphengine/src/common/graph/libgraph.so
+                ${CMAKE_BINARY_DIR}/graphengine/metadef/graph/libgraph.so
-                ${CMAKE_SOURCE_DIR}/build/graphengine/libc_sec.so
+                ${CMAKE_SOURCE_DIR}/build/graphengine/c_sec/lib/libc_sec.so
                 ${LIBEVENT_LIB_LIST}
             DESTINATION ${INSTALL_LIB_DIR}
             COMPONENT mindspore

graphengine

@@ -1 +1 @@
-Subproject commit ce23341ee2c7e48b39f227b293d1aa7f54effd34
+Subproject commit 20a0326976db65ca01f43ae4ccdd85677faaeb5e

mindspore/ccsrc/runtime/device/ascend/executor/ai_core_dynamic_kernel.cc

@@ -74,17 +74,14 @@ void AiCoreDynamicKernel::ParseCompileJson() {
   std::replace(compile_info_attr.begin(), compile_info_attr.end(), '\'', '\"');
   compile_info_attr = ReplaceInvalidJsonStr(compile_info_attr);
   MS_LOG(INFO) << "Get compile_info:" << compile_info_attr;
-
+  op_compile_info_.str = compile_info_attr;
-  try {
+  op_compile_info_.key = "";
-    compile_info_json_ = std::make_shared<nlohmann::json>(nlohmann::json::parse(compile_info_attr));
-  } catch (nlohmann::json::parse_error &e) {
-    MS_LOG(EXCEPTION) << "parse json failed, error:" << e.what();
-  }
 
   if (AnfAlgo::HasNodeAttr(kAttrFusionType, cnode_ptr_)) {
     auto fusion_type = AnfAlgo::GetNodeAttr<std::string>(cnode_ptr_, kAttrFusionType);
     MS_LOG(INFO) << "Get fusion_type:" << fusion_type;
     (*compile_info_json_)["_pattern"] = fusion_type;
+    op_compile_info_.key = std::hash<std::string>{}(fusion_type);
   }
 }
 
@@ -132,8 +129,8 @@ void AiCoreDynamicKernel::ComputeTiling() {
   MS_LOG(INFO) << "Start compute tiling of:" << cnode_ptr_->fullname_with_scope();
   optiling::OpRunInfo op_run_info;
 
-  OpTilingCalculater::GetInstance().CalculateTiling(NOT_NULL(cnode_ptr_), NOT_NULL(compile_info_json_),
+  OpTilingCalculater::GetInstance().CalculateTiling(NOT_NULL(cnode_ptr_), op_compile_info_, depend_tensor_map_,
-                                                    depend_tensor_map_, NOT_NULL(&op_run_info));
+                                                    NOT_NULL(&op_run_info));
   block_dim_ = op_run_info.block_dim;
   workspaces_size_ = op_run_info.workspaces;
   tiling_data_ = op_run_info.tiling_data.str();

mindspore/ccsrc/runtime/device/ascend/executor/ai_core_dynamic_kernel.h

@@ -24,6 +24,7 @@
 #include "nlohmann/json.hpp"
 #include "ir/tensor.h"
 #include "runtime/device/device_address.h"
+#include "register/op_tiling.h"
 #include "mindspore/ccsrc/runtime/device/executor/dynamic_kernel.h"
 
 namespace mindspore {
@@ -60,6 +61,7 @@ class AiCoreDynamicKernel : public DynamicKernel {
   std::vector<int64_t> workspaces_size_;
   std::vector<DeviceAddressPtr> workspace_addr_;
   std::shared_ptr<nlohmann::json> compile_info_json_;
+  optiling::OpCompileInfo op_compile_info_{};
 
   void ComputeTiling();
   bool CopyTilingToDevice();

mindspore/ccsrc/runtime/device/ascend/executor/tiling/op_tiling_calculater.cc

@@ -25,6 +25,7 @@
 #include "runtime/device/ascend/ge_types_convert.h"
 #include "utils/utils.h"
 #include "external/graph/tensor.h"
+#include "external/register/op_tiling_registry.h"
 
 namespace mindspore {
 namespace device {
@@ -136,7 +137,7 @@ void FeedTeOpConstTensor(const NotNull<CNodePtr> &cnode, const std::map<uint32_t
 
 void OpTilingCalculater::Init() {
   MS_LOG(INFO) << "Start init OpTilingCalculater";
-  tiling_func_map_ = optiling::OpTilingInterf::RegisteredOpInterf();
+  tiling_func_map_ = optiling::OpTilingRegistryInterf::RegisteredOpInterf();
   MS_LOG(INFO) << "tiling_func_map_ size:" << tiling_func_map_.size();
   for (const auto &iter : tiling_func_map_) {
     MS_LOG(INFO) << "Regist tiling func:" << iter.first;
@@ -157,8 +158,7 @@ std::string GetRealOpType(const std::string &op_type) {
   return iter->second;
 }
 
-void OpTilingCalculater::CalculateTiling(const NotNull<CNodePtr> &cnode,
+void OpTilingCalculater::CalculateTiling(const NotNull<CNodePtr> &cnode, const optiling::OpCompileInfo &op_compile_info,
-                                         const NotNull<std::shared_ptr<nlohmann::json>> &compile_info_json,
                                          const std::map<uint32_t, tensor::TensorPtr> &depend_tensor_map,
                                          NotNull<optiling::OpRunInfo *> op_run_info) {
   optiling::TeOpParas op_param;
@@ -181,7 +181,7 @@ void OpTilingCalculater::CalculateTiling(const NotNull<CNodePtr> &cnode,
   MS_LOG(INFO) << "Get tiling func:" << iter->first;
 
   if (iter != tiling_func_map_.end()) {
-    bool ret = (iter->second)(op_type, op_param, *compile_info_json.get(), *op_run_info);
+    bool ret = (iter->second)(op_param, op_compile_info, *op_run_info);
     if (!ret) {
       MS_LOG(EXCEPTION) << "Calculate tiling failed";
     }

mindspore/ccsrc/runtime/device/ascend/executor/tiling/op_tiling_calculater.h

@@ -37,8 +37,7 @@ class OpTilingCalculater {
   }
 
   void Init();
-  void CalculateTiling(const NotNull<CNodePtr> &cnode,
+  void CalculateTiling(const NotNull<CNodePtr> &cnode, const optiling::OpCompileInfo &op_compile_info,
-                       const NotNull<std::shared_ptr<nlohmann::json>> &compile_info_json,
                        const std::map<uint32_t, tensor::TensorPtr> &depend_tensor_map,
                        NotNull<optiling::OpRunInfo *> op_run_info);
 

mindspore/ccsrc/runtime/hccl_adapter/hccl_adapter.h

@@ -21,7 +21,7 @@
 #include <vector>
 #include <memory>
 #include "mindspore/core/ir/anf.h"
-#include "external/hccl/hccl_types.h"
+#include "hccl/hccl_types.h"
 
 #define MS_API __attribute__((visibility("default")))
 

tests/st/model_zoo_tests/transformer/test_transformer.py

@@ -164,10 +164,10 @@ def test_transformer():
 
     # assertion occurs while the loss value, overflow state or loss_scale value is wrong
     loss_value = np.array(callback.loss_list)
-    assert np.allclose(loss_value[0], 11.241624, 0, 0.000005)
+    assert np.allclose(loss_value[0], 11.241604, 0, 0.000005)
 
-    expect_loss_value = [11.241624, 11.243232, 11.217465, 11.204196, 11.2138195,
+    expect_loss_value = [11.241604, 11.243231, 11.217458, 11.204156, 11.213805,
-                         11.215386, 11.19053, 11.150403, 11.191858, 11.160057]
+                         11.215374, 11.19065, 11.150393, 11.191824, 11.160044]
     print("loss value: {}".format(loss_value))
     assert np.allclose(loss_value[0:10], expect_loss_value, 0, 0.0005)
 

tests/ut/cpp/runtest.sh

@@ -15,7 +15,7 @@
 # ============================================================================
 
 set -e
-BASEPATH=$(cd $(dirname $0); pwd)
+BASEPATH=$(cd "$(dirname "$0")"; pwd)
 PROJECT_PATH=${BASEPATH}/../../..
 if [ $BUILD_PATH ];then
 	echo "BUILD_PATH = $BUILD_PATH"
@@ -26,7 +26,9 @@ fi
 cd ${BUILD_PATH}/mindspore/tests/ut/cpp
 
 
-export LD_LIBRARY_PATH=${BUILD_PATH}/mindspore/googletest/googlemock/gtest:${PROJECT_PATH}/mindspore:${PROJECT_PATH}/mindspore/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=${BUILD_PATH}/mindspore/googletest/googlemock/gtest:${PROJECT_PATH}/mindspore:\
+${PROJECT_PATH}/mindspore/lib:${PROJECT_PATH}/graphengine/third_party/prebuild/x86_64:\
+${PROJECT_PATH}/graphengine/third_party/prebuild/aarch64:${LD_LIBRARY_PATH}
 export PYTHONPATH=${PROJECT_PATH}/tests/ut/cpp/python_input:$PYTHONPATH:${PROJECT_PATH}
 export GLOG_v=2
 export GC_COLLECT_IN_CELL=1

tests/ut/cpp/stub/dynamic_shape/dynamic_shape_stub.cc

@@ -53,8 +53,7 @@ bool HcclExecutorManager::Initialize() { return true; }
 bool HcclExecutorManager::Finalize() { return true; }
 
 void OpTilingCalculater::Init() {}
-void OpTilingCalculater::CalculateTiling(const NotNull<CNodePtr> &cnode,
+void OpTilingCalculater::CalculateTiling(const NotNull<CNodePtr> &cnode, const optiling::OpCompileInfo &op_compile_info,
-                                         const NotNull<std::shared_ptr<nlohmann::json>> &compile_info_json,
                      const std::map<uint32_t, tensor::TensorPtr> &depend_tensor_map,
                      NotNull<optiling::OpRunInfo *> op_run_info) {}
 }  // namespace ascend