fix(mgb/core): fix memory management release cambricon var issue

GitOrigin-RevId: abf881978c8fd85b3cfa823fea13045ff06fb88e

CMakeLists.txt

@@ -513,6 +513,7 @@ if(MGE_WITH_ATLAS)
     set(MGB_ATLAS ${MGE_WITH_ATLAS})
 endif()
 
+
 find_program(CCACHE_BIN ccache)
 if(CCACHE_BIN)
     set(CMAKE_CXX_COMPILER_LAUNCHER ${CCACHE_BIN})
@@ -688,6 +689,7 @@ if(MGE_ARCH STREQUAL "aarch64")
     set(MEGDNN_AARCH64 1)
     set(MEGDNN_64_BIT 1)
     set(MARCH "-march=armv8-a")
+    set(MGB_AARCH64 1)
     if(MGE_ARMV8_2_FEATURE_FP16)
         message(STATUS "Enable fp16 feature support in armv8.2")
         if(NOT ${MGE_DISABLE_FLOAT16})

src/core/impl/comp_node_env.cpp

@@ -35,6 +35,7 @@
 #include "megcore_atlas.h"
 #endif
 
+
 using namespace mgb;
 
 /* =================== MegDNNHandle =================== */
@@ -101,6 +102,7 @@ MegDNNHandle::MegDNNHandle(const CompNodeEnv& env) {
     }
 #endif
 
+
     if (env.property().type == CompNode::DeviceType::CPU) {
         megcoreCreateDeviceHandle(&m_dev_hdl, megcorePlatformCPU);
         megcoreCreateComputingHandleWithCPUDispatcher(&m_comp_hdl, m_dev_hdl,
@@ -254,6 +256,7 @@ void CompNodeEnv::init_atlas(CompNode comp_node, const AtlasEnv& env) {
 #endif
 
 
+
 #if MGB_ROCM
 
 void mgb::_on_hip_error(const char* expr, hipError_t err, const char* file,

src/core/impl/exception.cpp

@@ -77,6 +77,7 @@ AtlasError::AtlasError(const std::string &msg):
 }
 
 
+
 ROCmError::ROCmError(const std::string &msg):
     SystemError(msg)
 {

src/core/impl/graph/var_node_mem_mgr.cpp

@@ -125,7 +125,7 @@ StaticDeviceMemoryManager::make_default_impl() {
 #endif  // MGB_THREAD_SAFE
 
 /* ==================== AsyncVarReleaser ==================== */
-#if MGB_CUDA || MGB_ATLAS
+#if MGB_CUDA || MGB_ATLAS || MGB_CAMBRICON 
 class VarNodeMemManager::AsyncVarReleaser {
     struct WaiterParam {
         CompNode cn;
@@ -245,10 +245,10 @@ bool VarNodeMemManager::ImpureMemPlanManager::check_need_realloc() {
 }
 
 /* ==================== VarNodeMemManager ==================== */
-VarNodeMemManager::VarNodeMemManager(ComputingGraphImpl *graph):
-    m_owner_graph(graph),
+VarNodeMemManager::VarNodeMemManager(ComputingGraphImpl* graph)
+        : m_owner_graph(graph),
           m_seq_mem_opt(graph)
-#if MGB_CUDA || MGB_ATLAS
+#if MGB_CUDA || MGB_ATLAS || MGB_CAMBRICON 
           ,m_asyn_var_releaser(new AsyncVarReleaser)
 #endif
 {
@@ -256,7 +256,7 @@ VarNodeMemManager::VarNodeMemManager(ComputingGraphImpl *graph):
         MGB_MARK_USED_VAR(ev);
         // async release is only used for sync between multiple comp nodes, and
         // does not wait for device to finish
-#if MGB_CUDA || MGB_ATLAS
+#if MGB_CUDA || MGB_ATLAS || MGB_CAMBRICON 
         m_asyn_var_releaser->wait_release_finish();
 #endif
         m_cpu_async_release_barrier.wait_zero();
@@ -297,7 +297,8 @@ VarNodeMemManager::VarNodeMemManager(ComputingGraphImpl *graph):
     graph->event().register_receiver_permanent<event::CompSeqExecError>(
             on_comp_seq_error);
 
-#if MGB_ENABLE_VAR_DEV_MEM_DEFRAGMENTER && (MGB_CUDA || MGB_ATLAS)
+#if MGB_ENABLE_VAR_DEV_MEM_DEFRAGMENTER &&                                   \
+        (MGB_CUDA || MGB_ATLAS || MGB_CAMBRICON ) 
     auto on_mem_defrag_start = [this](const event::BeforeMemDefrag&) {
         m_asyn_var_releaser->wait_release_finish();
     };
@@ -1448,6 +1449,13 @@ void VarNodeMemManager::decr_var_mem_refcnt(
                 m_asyn_var_releaser->add(dispatch_cn, var);
                 break;
             }
+#endif
+#if MGB_CAMBRICON
+        case DT::CAMBRICON:
+            {
+                m_asyn_var_releaser->add(dispatch_cn, var);
+                break;
+            }
 #endif
         default:
             mgb_throw(MegBrainError,

src/core/impl/graph/var_node_mem_mgr.h

@@ -446,7 +446,7 @@ class VarNodeMemManager {
         SyncableCounter m_cpu_async_release_barrier;
 
 
-#if MGB_CUDA || MGB_ATLAS
+#if MGB_CUDA || MGB_ATLAS || MGB_CAMBRICON 
         //! release dynamic var on after compnode event finishes
         class AsyncVarReleaser;
         std::unique_ptr<AsyncVarReleaser> m_asyn_var_releaser;

src/core/include/megbrain/comp_node_env.h

@@ -90,6 +90,7 @@
 #endif // MGB_ATLAS
 
 
+
 #if MGB_ROCM
 #include "hcc_detail/hcc_defs_prologue.h"
 #include "megcore_rocm.h"
@@ -194,6 +195,7 @@ namespace mgb {
                                   const char* file, const char* func, int line);
 #endif
 
+
 #if MGB_CUDA
 [[noreturn]] void _on_cuda_error(const char* expr, cudaError_t err,
                                  const char* file, const char* func, int line);
@@ -325,6 +327,7 @@ public:
         }
 #endif
 
+
     }
 
     /*!
@@ -426,6 +429,8 @@ public:
     void init_atlas(CompNode comp_node, const AtlasEnv& env);
 #endif
 
+
+
 #if MGB_ROCM
     struct ROCmEnv {
         int device = -1;
@@ -485,9 +490,7 @@ public:
         };
         static InitStatus init_status;
 
-        static void init() {
-            init_status.init();
-        }
+        static void init() { init_status.init(); }
 
         void activate() const {
             init();

src/core/test/comp_node.cpp

@@ -62,6 +62,7 @@ TEST(TestCompNode, Parse) {
     ASSERT_EQ(L::parse("multithread:default:2"),
               make_lc(D::MULTITHREAD, L::DEVICE_MULTITHREAD_DEFAULT, 2));
 
+
     ASSERT_THROW(L::parse("apu"), MegBrainError);
     ASSERT_THROW(L::parse("fpgbx"), MegBrainError);
     ASSERT_THROW(L::parse("cab0"), MegBrainError);
@@ -149,6 +150,7 @@ TEST(TestCompNode, Load) {
     auto atlas1 = CompNode::load("atlas1");
     ASSERT_NE(atlas0, atlas1);
 #endif
+
 }
 
 TEST(TestCompNode, FreeAfterFinalize) {
@@ -762,6 +764,7 @@ TEST(TestCompNodeAtlas, D2DCopy) {
 }
 #endif
 
+
 namespace {
 class CompNodeDepedentObjectInst final : public CompNodeDepedentObject {
     int *m_dst, *m_timer;

src/megbrain_build_config.h.in

@@ -33,7 +33,6 @@
 #cmakedefine01 MGB_ENABLE_OPR_MM
 #cmakedefine01 MGB_ENABLE_FBS_SERIALIZATION
 #cmakedefine01 MGB_IS_DEV
-
 // DNN related flags
 // Platform macro's
 #cmakedefine01 MEGDNN_WITH_CUDA