GeTensor aligned addr & zero copy support

ge/CMakeLists.txt

@@ -125,7 +125,7 @@ set(TRAIN_SRC_LIST
     "graph/manager/graph_var_manager.cc"
     "graph/manager/host_mem_manager.cc"
     "graph/manager/rdma_pool_allocator.cc"
-    $<$<STREQUAL:${ENABLE_OPEN_SRC},FALSE>:graph/manager/host_mem_allocator.cc>
+    $<$<NOT:$<STREQUAL:${ENABLE_OPEN_SRC},True>>:graph/manager/host_mem_allocator.cc>
     "graph/manager/memory_api.cc"
     "graph/manager/model_manager/event_manager.cc"
     "graph/manager/trans_var_data_utils.cc"
@@ -167,7 +167,7 @@ set(TRAIN_SRC_LIST
     "graph/passes/hccl_group_pass.cc"
     "graph/passes/enter_pass.cc"
     "graph/passes/assign_pass.cc"
-    $<$<STREQUAL:${ENABLE_OPEN_SRC},FALSE>:graph/passes/inplace_support_check_pass.cc>
+    $<$<NOT:$<STREQUAL:${ENABLE_OPEN_SRC},True>>:graph/passes/inplace_support_check_pass.cc>
     "graph/passes/flow_ctrl_pass.cc"
     "graph/passes/global_step_insert_pass.cc"
     "host_kernels/transpose_kernel.cc"
@@ -403,7 +403,7 @@ set(INFER_SRC_LIST
     "graph/manager/graph_var_manager.cc"
     "graph/manager/host_mem_manager.cc"
     "graph/manager/rdma_pool_allocator.cc"
-    $<$<STREQUAL:${ENABLE_OPEN_SRC},FALSE>:graph/manager/host_mem_allocator.cc>
+    $<$<NOT:$<STREQUAL:${ENABLE_OPEN_SRC},True>>:graph/manager/host_mem_allocator.cc>
     "graph/manager/graph_mem_allocator.cc"
     "graph/manager/graph_caching_allocator.cc"
     "model/ge_model.cc"
@@ -525,7 +525,7 @@ set(INFER_SRC_LIST
     "graph/passes/for_pass.cc"
     "graph/passes/enter_pass.cc"
     "graph/passes/assign_pass.cc"
-    $<$<STREQUAL:${ENABLE_OPEN_SRC},FALSE>:graph/passes/inplace_support_check_pass.cc>
+    $<$<NOT:$<STREQUAL:${ENABLE_OPEN_SRC},True>>:graph/passes/inplace_support_check_pass.cc>
     "graph/passes/addn_pass.cc"
     "graph/passes/common_subexpression_elimination_pass.cc"
     "graph/passes/remove_same_const_pass.cc"
@@ -624,6 +624,7 @@ target_compile_definitions(ge_runner PRIVATE
     FMK_SUPPORT_DUMP
     DAVINCI_CLOUD
     google=ascend_private
+    $<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC>
 )
 
 target_compile_options(ge_runner PRIVATE
@@ -691,6 +692,7 @@ target_compile_definitions(ge_compiler PRIVATE
     FMK_HOST_INFER
     COMPILE_OMG_PACKAGE
     google=ascend_private
+    $<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC>
 )
 
 target_compile_options(ge_compiler PRIVATE

ge/executor/CMakeLists.txt

@@ -28,7 +28,7 @@ set(SRC_LIST
     "../graph/manager/trans_var_data_utils.cc"
     "../graph/manager/util/debug.cc"
     "../graph/manager/rdma_pool_allocator.cc"
-    $<$<STREQUAL:${ENABLE_OPEN_SRC},FALSE>:../graph/manager/host_mem_allocator.cc>
+    $<$<NOT:$<STREQUAL:${ENABLE_OPEN_SRC},True>>:../graph/manager/host_mem_allocator.cc>
     "../hybrid/node_executor/aicpu/aicpu_ext_info.cc"
     "../model/ge_model.cc"
     "../model/ge_root_model.cc"
@@ -175,6 +175,7 @@ target_compile_definitions(ge_executor PRIVATE
     $<IF:$<STREQUAL:${TARGET_SYSTEM_NAME},Windows>,OS_TYPE=WIN,OS_TYPE=0>
     $<$<STREQUAL:${TARGET_SYSTEM_NAME},Windows>:SECUREC_USING_STD_SECURE_LIB=0 NOMINMAX>
     LOG_CPP
+    $<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC>
 )
 
 target_include_directories(ge_executor PRIVATE
@@ -217,6 +218,7 @@ target_compile_definitions(ge_executor_shared PRIVATE
     PROTOBUF_INLINE_NOT_IN_HEADERS=0
     DAVINCI_SUPPORT_PROFILING
     google=ascend_private
+    $<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC>
 )
 
 target_include_directories(ge_executor_shared PRIVATE

ge/ge_local_engine/engine/host_cpu_engine.cc

@@ -26,7 +26,7 @@
 #include "common/math/math_util.h"
 
 namespace {
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 #define CREATE_OUTPUT_CASE(DTYPE, TYPE)                                                                                \
   case (DTYPE): {                                                                                                      \
     GeTensorPtr ge_tensor = nullptr;                                                                                   \

ge/graph/manager/graph_manager.cc

@@ -38,7 +38,7 @@
 #include "graph/partition/stage_partition.h"
 #include "graph/passes/addn_pass.h"
 #include "graph/passes/bitcast_pass.h"
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 #include "graph/passes/assign_pass.h"
 #include "graph/passes/inplace_support_check_pass.h"
 #endif
@@ -2241,7 +2241,7 @@ Status GraphManager::OptimizeStage2(ge::ComputeGraphPtr &compute_graph) {
   ReshapeRemovePass reshape_remove_pass;
   CondRemovePass condition_remove_pass;
   BitcastPass bitcast_pass;
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   AssignPass assign_pass;
   InplaceSupportCheckPass inplace_support_check_pass;
 #endif
@@ -2249,7 +2249,7 @@ Status GraphManager::OptimizeStage2(ge::ComputeGraphPtr &compute_graph) {
   names_to_passes.emplace_back("ReshapeRemovePass", &reshape_remove_pass);
   names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass);
   names_to_passes.emplace_back("BitcastPass", &bitcast_pass);
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   if (GetContext().GetHostExecFlag()) {
     names_to_passes.emplace_back("AssignPass", &assign_pass);
     names_to_passes.emplace_back("InplaceSupportCheckPass", &inplace_support_check_pass);

ge/graph/manager/graph_mem_allocator.cc

@@ -19,7 +19,7 @@
 #include <string>
 #include "graph/manager/graph_caching_allocator.h"
 #include "graph/manager/rdma_pool_allocator.h"
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 #include "graph/manager/host_mem_allocator.h"
 #endif
 namespace ge {
@@ -192,7 +192,7 @@ Status MemManager::Initialize(const std::vector<rtMemType_t> &memory_type) {
     GELOGE(ge::INTERNAL_ERROR, "Create RdmaAllocator failed.");
     return ge::INTERNAL_ERROR;
   }
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   if (InitAllocator(memory_type, host_allocator_map_) != SUCCESS) {
     GELOGE(ge::INTERNAL_ERROR, "Create HostMemAllocator failed.");
     return ge::INTERNAL_ERROR;
@@ -219,7 +219,7 @@ void MemManager::Finalize() noexcept {
   // caching and rdma allocator use memory allocator, so finalize them first
   FinalizeAllocatorMap(caching_allocator_map_);
   FinalizeAllocatorMap(rdma_allocator_map_);
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   FinalizeAllocatorMap(host_allocator_map_);
 #endif
   FinalizeAllocatorMap(memory_allocator_map_);
@@ -250,7 +250,7 @@ CachingAllocator &MemManager::CachingInstance(rtMemType_t memory_type) {
 RdmaPoolAllocator &MemManager::RdmaPoolInstance(rtMemType_t memory_type) {
   return Instance().GetAllocator(memory_type, rdma_allocator_map_);
 }
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 HostMemAllocator &MemManager::HostMemInstance(rtMemType_t memory_type) {
   return Instance().GetAllocator(memory_type, host_allocator_map_);
 }

ge/graph/manager/graph_mem_allocator.h

@@ -139,7 +139,7 @@ class MemoryAllocator {
 using MemoryAllocatorPtr = std::shared_ptr<MemoryAllocator>;
 class CachingAllocator;
 class RdmaPoolAllocator;
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 class HostMemAllocator;
 #endif
 class MemManager {
@@ -150,7 +150,7 @@ class MemManager {
   static MemoryAllocator *Instance(rtMemType_t memory_type);
   CachingAllocator &CachingInstance(rtMemType_t memory_type);
   RdmaPoolAllocator &RdmaPoolInstance(rtMemType_t memory_type);
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   HostMemAllocator &HostMemInstance(rtMemType_t memory_type);
 #endif
   MemManager(const MemManager &) = delete;
@@ -240,7 +240,7 @@ class MemManager {
   std::map<rtMemType_t, MemoryAllocator *> memory_allocator_map_;
   std::map<rtMemType_t, CachingAllocator *> caching_allocator_map_;
   std::map<rtMemType_t, RdmaPoolAllocator *> rdma_allocator_map_;
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   std::map<rtMemType_t, HostMemAllocator *> host_allocator_map_;
 #endif
   std::recursive_mutex allocator_mutex_;

ge/graph/manager/host_mem_manager.cc

@@ -43,7 +43,7 @@ Status SharedMemAllocator::Allocate(SharedMemInfo &mem_info) {
     return GE_GRAPH_MEMORY_ALLOC_FAILED;
   }
   mem_info.fd = output_para.fd;
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   mem_info.host_aligned_ptr = AlignedPtr::BuildAlignedPtr(mem_info.mem_size,
                                                           [&output_para](std::unique_ptr<uint8_t[], deleter> &ptr) {
                                                               GELOGD("set aligned_ptr, addr=%p", output_para.ptr);
@@ -62,7 +62,7 @@ Status SharedMemAllocator::Allocate(SharedMemInfo &mem_info) {
 
 Status SharedMemAllocator::DeAllocate(SharedMemInfo &mem_info) {
   GELOGD("SharedMemAllocator::DeAllocate");
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   rtFreeHostSharedMemoryIn free_para = {mem_info.shm_name.c_str(), mem_info.mem_size, mem_info.fd,
                                         mem_info.host_aligned_ptr->MutableGet(), mem_info.device_address};
 #else

ge/graph/manager/host_mem_manager.h

@@ -42,7 +42,7 @@ struct SharedMemInfo {
   uint64_t mem_size = 0;
   int fd = 0;
   uint8_t *device_address = nullptr;
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   std::shared_ptr<AlignedPtr> host_aligned_ptr = nullptr;
 #else
   uint8_t *host_address = nullptr;

ge/graph/passes/assign_pass.cc

@@ -26,7 +26,7 @@ const int32_t kAssignValueInputIndex = 1;
 }
 
 namespace ge {
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 Status AssignPass::Run(NodePtr &node) {
   GELOGD("AssignPass running");
 

ge/graph/passes/assign_pass.h

@@ -25,7 +25,7 @@ class AssignPass : public BaseNodePass {
   Status Run(NodePtr &node) override;
 
  private:
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   ///
   /// @brief Optimize for assign_node
   /// @param [in] assign_node

ge/graph/passes/constant_fuse_same_pass.cc

@@ -115,7 +115,7 @@ void ConstantFuseSamePass::GetFuseConstNodes(ComputeGraphPtr &graph,
              TypeUtils::DataTypeToSerialString(data_type).c_str());
       continue;
     }
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
     if ((type_size != 0) && (weight->MutableData().GetAlignedPtr() == nullptr)) {
       GELOGW("aligned_ptr is null while size is not 0");
       continue;
@@ -125,7 +125,7 @@ void ConstantFuseSamePass::GetFuseConstNodes(ComputeGraphPtr &graph,
 
     SameConstKey map_key;
     map_key.data_size = type_size;
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
     map_key.aligned_ptr = weight->MutableData().GetAlignedPtr();
 #else
     map_key.data = weight->GetData().GetData();

ge/graph/passes/constant_fuse_same_pass.h

@@ -21,7 +21,7 @@
 #include <set>
 #include <utility>
 #include <vector>
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 #include "graph/aligned_ptr.h"
 #endif
 #include "graph/types.h"
@@ -30,7 +30,7 @@
 namespace ge {
 struct SameConstKey {
   int data_size;
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
   std::shared_ptr<AlignedPtr> aligned_ptr;
 #else
   const uint8_t *data;
@@ -44,7 +44,7 @@ struct SameConstKey {
     if (data_size != key.data_size) {
       return data_size < key.data_size;
     }
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
     if (data_size != 0) {
       int ret = memcmp(aligned_ptr->Get(), key.aligned_ptr->Get(), data_size);
       if (ret != 0) {

ge/graph/preprocess/graph_preprocess.cc

@@ -37,7 +37,7 @@
 #include "graph/passes/addn_pass.h"
 #include "graph/passes/aicpu_constant_folding_pass.h"
 #include "graph/passes/assert_pass.h"
-#if (ENABLE_OPEN_SRC == True)
+#ifdef ONLY_COMPILE_OPEN_SRC
 #include "graph/passes/assign_pass.h"
 #endif
 #include "graph/passes/common_subexpression_elimination_pass.h"
@@ -1700,7 +1700,7 @@ Status GraphPrepare::PrepareOptimize() {
   VarIsInitializedOpPass var_is_initialized_pass;
   ParallelConcatStartOpPass parallel_concat_start_op_pass;
   IdentityPass identity_pass(false);
-#if (ENABLE_OPEN_SRC == True)
+#ifdef ONLY_COMPILE_OPEN_SRC
   AssignPass assign_pass;
 #endif
   SnapshotPass snapshot_pass;
@@ -1717,7 +1717,7 @@ Status GraphPrepare::PrepareOptimize() {
   names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass);
   names_to_passes.emplace_back("ParallelConcatStartOpPass", &parallel_concat_start_op_pass);
   names_to_passes.emplace_back("IdentityPass", &identity_pass);
-#if (ENABLE_OPEN_SRC == True)
+#ifdef ONLY_COMPILE_OPEN_SRC
   if (GetContext().GetHostExecFlag()) {
     names_to_passes.emplace_back("AssignPass", &assign_pass);
   }

ge/hybrid/common/npu_memory_allocator.cc

@@ -20,7 +20,7 @@
 #include "graph/manager/graph_caching_allocator.h"
 #include "graph/manager/graph_mem_allocator.h"
 #include "graph/manager/rdma_pool_allocator.h"
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 #include "graph/manager/host_mem_allocator.h"
 #endif
 
@@ -67,7 +67,7 @@ void *NpuMemoryAllocator::Allocate(std::size_t size, AllocationAttr *attr) {
   if (mem_type == RDMA_HBM) {
     buffer = MemManager::Instance().RdmaPoolInstance(RT_MEMORY_HBM).Malloc(allocate_size, device_id_);
   } else if (mem_type == HOST_DDR) {
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
     buffer = MemManager::Instance().HostMemInstance(RT_MEMORY_HBM).Malloc(allocate_size);
 #else
     buffer = malloc(allocate_size);
@@ -108,7 +108,7 @@ void NpuMemoryAllocator::Deallocate(void *data, MemStorageType mem_type) {
     if (mem_type == RDMA_HBM) {
       MemManager::Instance().RdmaPoolInstance(RT_MEMORY_HBM).Free(reinterpret_cast<uint8_t *>(data), device_id_);
     } else if (mem_type == HOST_DDR) {
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
       MemManager::Instance().HostMemInstance(RT_MEMORY_HBM).Free(data);
 #else
       free(data);

ge/hybrid/model/hybrid_model_builder.cc

@@ -24,7 +24,7 @@
 #include "graph/manager/graph_var_manager.h"
 #include "graph/manager/host_mem_manager.h"
 #include "graph/manager/trans_var_data_utils.h"
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 #include "graph/manager/graph_mem_allocator.h"
 #include "graph/manager/host_mem_allocator.h"
 #endif
@@ -853,7 +853,7 @@ Status HybridModelBuilder::InitConstantOps() {
 
     std::unique_ptr<TensorValue> var_tensor;
     if (GetContext().GetHostExecFlag()) {
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
       GE_CHECK_NOTNULL(ge_tensor);
       // Address for eigen kernel should be aligned with 16 bytes
       // Tensors return by api GetWeights share data with proto, whose addr is not confirmed to be aligned
@@ -925,7 +925,7 @@ Status HybridModelBuilder::InitVariableTensors() {
       GELOGE(GE_GRAPH_MALLOC_FAILED, "Host variable [%s] malloc failed.", it.first.c_str());
       return GE_GRAPH_MALLOC_FAILED;
     }
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
     if (MemManager::Instance().HostMemInstance(RT_MEMORY_HBM).Malloc(mem_info.host_aligned_ptr,
                                                                      tensor_size) == nullptr) {
       GELOGE(MEMALLOC_FAILED, "Malloc host memory for an existed GeTensor failed.");

ge/hybrid/node_executor/host_cpu/host_cpu_node_executor.cc

@@ -18,7 +18,7 @@
 #include "hybrid/node_executor/host_cpu/kernel_factory.h"
 #include "graph/passes/folding_pass.h"
 #include "hybrid/model/hybrid_model.h"
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
 #include "graph/manager/graph_mem_allocator.h"
 #include "graph/manager/host_mem_allocator.h"
 #endif
@@ -54,7 +54,7 @@ Status CpuKernelNodeTask::Execute(TaskContext &context) {
     auto input_desc_ptr = context.GetInputDesc(i);
     GE_CHECK_NOTNULL(input_desc_ptr);
     const auto &input_desc = *input_desc_ptr;
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
     auto tensor = context.GetInput(i);
     GE_CHECK_NOTNULL(tensor);
     auto item = MemManager::Instance().HostMemInstance(RT_MEMORY_HBM).GetAlignedPtr(tensor->GetData());
@@ -84,7 +84,7 @@ Status CpuKernelNodeTask::Execute(TaskContext &context) {
     }
     auto tensor = context.GetOutput(i);
     GE_CHECK_NOTNULL(tensor);
-#if (ENABLE_OPEN_SRC != True)
+#ifndef ONLY_COMPILE_OPEN_SRC
     auto item = MemManager::Instance().HostMemInstance(RT_MEMORY_HBM).GetAlignedPtr(tensor->GetData());
     GE_CHECK_NOTNULL(item.second);
     auto out_tensor = MakeShared<GeTensor>(output_desc, item.second, item.first);