diff --git a/ge/CMakeLists.txt b/ge/CMakeLists.txt index 8d9edb65..e94258ac 100755 --- a/ge/CMakeLists.txt +++ b/ge/CMakeLists.txt @@ -124,7 +124,7 @@ set(TRAIN_SRC_LIST "graph/manager/graph_var_manager.cc" "graph/manager/host_mem_manager.cc" "graph/manager/rdma_pool_allocator.cc" - $<$>:graph/manager/host_mem_allocator.cc> + "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" @@ -166,7 +166,7 @@ set(TRAIN_SRC_LIST "graph/passes/hccl_group_pass.cc" "graph/passes/enter_pass.cc" "graph/passes/assign_remove_pass.cc" - $<$>:graph/passes/inplace_support_check_pass.cc> + "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" @@ -409,7 +409,7 @@ set(INFER_SRC_LIST "graph/manager/graph_var_manager.cc" "graph/manager/host_mem_manager.cc" "graph/manager/rdma_pool_allocator.cc" - $<$>:graph/manager/host_mem_allocator.cc> + "graph/manager/host_mem_allocator.cc" "graph/manager/graph_mem_allocator.cc" "graph/manager/graph_caching_allocator.cc" "model/ge_model.cc" @@ -531,7 +531,7 @@ set(INFER_SRC_LIST "graph/passes/for_pass.cc" "graph/passes/enter_pass.cc" "graph/passes/assign_remove_pass.cc" - $<$>:graph/passes/inplace_support_check_pass.cc> + "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" diff --git a/ge/executor/CMakeLists.txt b/ge/executor/CMakeLists.txt index 4ca18864..ac4d4ebd 100644 --- a/ge/executor/CMakeLists.txt +++ b/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" - $<$>:../graph/manager/host_mem_allocator.cc> + "../graph/manager/host_mem_allocator.cc" "../hybrid/node_executor/aicpu/aicpu_ext_info.cc" "../model/ge_model.cc" "../model/ge_root_model.cc" diff --git a/ge/ge_local_engine/engine/host_cpu_engine.cc b/ge/ge_local_engine/engine/host_cpu_engine.cc index 0f46b4cb..35ecfb2d 100755 --- a/ge/ge_local_engine/engine/host_cpu_engine.cc +++ b/ge/ge_local_engine/engine/host_cpu_engine.cc @@ -26,7 +26,6 @@ #include "common/math/math_util.h" namespace { -#ifndef ONLY_COMPILE_OPEN_SRC #define CREATE_OUTPUT_CASE(DTYPE, TYPE) \ case (DTYPE): { \ GeTensorPtr ge_tensor = nullptr; \ @@ -50,43 +49,6 @@ namespace { named_outputs.emplace(tensor_name, tensor); \ break; \ } -#else -#define CREATE_OUTPUT_CASE(DTYPE, TYPE) \ - case (DTYPE): { \ - GeTensorPtr ge_tensor = nullptr; \ - if (need_create_flag) { \ - GELOGI("node:%s allocate output %zu start, size=%lld", op_desc->GetName().c_str(), i, data_num * sizeof(TYPE)); \ - std::unique_ptr buf(new (std::nothrow) TYPE[data_num]()); \ - if (buf == nullptr) { \ - GELOGE(MEMALLOC_FAILED, "New sizeof(T) * data_num(%zu) memory failed", \ - static_cast(sizeof(TYPE) * data_num)); \ - return MEMALLOC_FAILED; \ - } \ - ge_tensor = MakeShared(out_desc); \ - GE_CHECK_NOTNULL(ge_tensor); \ - GELOGD("node:%s allocate output %zu success, size=%lld", op_desc->GetName().c_str(), i, data_num * sizeof(TYPE));\ - if (ge_tensor->SetData(reinterpret_cast(buf.get()), data_num * sizeof(TYPE)) != GRAPH_SUCCESS) { \ - GELOGE(MEMALLOC_FAILED, "Set data for output %zu of node %s failed.", i, op_desc->GetName().c_str()); \ - return MEMALLOC_FAILED; \ - } \ - ge_tensor->MutableTensorDesc().SetDataType(out_desc.GetDataType()); \ - ge_tensor->MutableTensorDesc().SetShape(out_desc.GetShape()); \ - outputs.emplace_back(ge_tensor); \ - } else { \ - ge_tensor = outputs[i]; \ - GE_CHECK_NOTNULL(ge_tensor); \ - GELOGD("node:%s existed output %zu", op_desc->GetName().c_str(), i); \ - } \ - auto tensor = TensorAdapter::AsTensor(*ge_tensor); \ - auto tensor_name = op_desc->GetOutputNameByIndex(i); \ - GE_RETURN_WITH_LOG_IF_TRUE(tensor_name.empty(), "Failed to get output name. node = %s, index = %zu", \ - op_desc->GetName().c_str(), i); \ - GELOGD("Successfully inserted output tensor. node = %s, index = %zu, output name = %s, addr = %p, size = %zu", \ - op_desc->GetName().c_str(), i, tensor_name.c_str(), tensor.GetData(), tensor.GetSize()); \ - named_outputs.emplace(tensor_name, tensor); \ - break; \ - } -#endif } namespace ge { diff --git a/ge/graph/manager/graph_manager.cc b/ge/graph/manager/graph_manager.cc index c4f91036..c0f084d8 100755 --- a/ge/graph/manager/graph_manager.cc +++ b/ge/graph/manager/graph_manager.cc @@ -38,10 +38,8 @@ #include "graph/partition/stage_partition.h" #include "graph/passes/addn_pass.h" #include "graph/passes/bitcast_pass.h" -#ifndef ONLY_COMPILE_OPEN_SRC #include "graph/passes/assign_remove_pass.h" #include "graph/passes/inplace_support_check_pass.h" -#endif #include "graph/passes/atomic_addr_clean_pass.h" #include "graph/passes/attach_stream_label_pass.h" #include "graph/passes/cast_remove_pass.h" @@ -2269,20 +2267,16 @@ Status GraphManager::OptimizeStage2(ge::ComputeGraphPtr &compute_graph) { ReshapeRemovePass reshape_remove_pass; CondRemovePass condition_remove_pass; BitcastPass bitcast_pass; -#ifndef ONLY_COMPILE_OPEN_SRC AssignRemovePass assign_remove_pass; InplaceSupportCheckPass inplace_support_check_pass; -#endif names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass); 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); -#ifndef ONLY_COMPILE_OPEN_SRC if (GetContext().GetHostExecFlag()) { names_to_passes.emplace_back("AssignRemovePass", &assign_remove_pass); names_to_passes.emplace_back("InplaceSupportCheckPass", &inplace_support_check_pass); } -#endif GE_TIMESTAMP_START(names_to_passes); ret = GEPass(compute_graph).Run(names_to_passes); GE_TIMESTAMP_END(names_to_passes, "OptimizeStage2::MergedGraphNameToPasses"); diff --git a/ge/graph/manager/graph_mem_allocator.cc b/ge/graph/manager/graph_mem_allocator.cc index 4e31d835..f3037299 100755 --- a/ge/graph/manager/graph_mem_allocator.cc +++ b/ge/graph/manager/graph_mem_allocator.cc @@ -19,9 +19,7 @@ #include #include "graph/manager/graph_caching_allocator.h" #include "graph/manager/rdma_pool_allocator.h" -#ifndef ONLY_COMPILE_OPEN_SRC #include "graph/manager/host_mem_allocator.h" -#endif namespace ge { void MemoryAllocator::Initialize(uint32_t device_id) { GELOGI("MemoryAllocator::Initialize"); @@ -192,12 +190,10 @@ Status MemManager::Initialize(const std::vector &memory_type) { GELOGE(ge::INTERNAL_ERROR, "Create RdmaAllocator failed."); return ge::INTERNAL_ERROR; } -#ifndef ONLY_COMPILE_OPEN_SRC if (InitAllocator(memory_type, host_allocator_map_) != SUCCESS) { GELOGE(ge::INTERNAL_ERROR, "Create HostMemAllocator failed."); return ge::INTERNAL_ERROR; } -#endif return SUCCESS; } @@ -219,9 +215,7 @@ void MemManager::Finalize() noexcept { // caching and rdma allocator use memory allocator, so finalize them first FinalizeAllocatorMap(caching_allocator_map_); FinalizeAllocatorMap(rdma_allocator_map_); -#ifndef ONLY_COMPILE_OPEN_SRC FinalizeAllocatorMap(host_allocator_map_); -#endif FinalizeAllocatorMap(memory_allocator_map_); } @@ -250,9 +244,7 @@ CachingAllocator &MemManager::CachingInstance(rtMemType_t memory_type) { RdmaPoolAllocator &MemManager::RdmaPoolInstance(rtMemType_t memory_type) { return Instance().GetAllocator(memory_type, rdma_allocator_map_); } -#ifndef ONLY_COMPILE_OPEN_SRC HostMemAllocator &MemManager::HostMemInstance(rtMemType_t memory_type) { return Instance().GetAllocator(memory_type, host_allocator_map_); } -#endif } // namespace ge diff --git a/ge/graph/manager/graph_mem_allocator.h b/ge/graph/manager/graph_mem_allocator.h index 6cdbd9b4..bd75dbb9 100644 --- a/ge/graph/manager/graph_mem_allocator.h +++ b/ge/graph/manager/graph_mem_allocator.h @@ -139,9 +139,7 @@ class MemoryAllocator { using MemoryAllocatorPtr = std::shared_ptr; class CachingAllocator; class RdmaPoolAllocator; -#ifndef ONLY_COMPILE_OPEN_SRC class HostMemAllocator; -#endif class MemManager { public: MemManager(); @@ -150,9 +148,7 @@ class MemManager { static MemoryAllocator *Instance(rtMemType_t memory_type); CachingAllocator &CachingInstance(rtMemType_t memory_type); RdmaPoolAllocator &RdmaPoolInstance(rtMemType_t memory_type); -#ifndef ONLY_COMPILE_OPEN_SRC HostMemAllocator &HostMemInstance(rtMemType_t memory_type); -#endif MemManager(const MemManager &) = delete; MemManager &operator=(const MemManager &) = delete; /// @@ -240,9 +236,7 @@ class MemManager { std::map memory_allocator_map_; std::map caching_allocator_map_; std::map rdma_allocator_map_; -#ifndef ONLY_COMPILE_OPEN_SRC std::map host_allocator_map_; -#endif std::recursive_mutex allocator_mutex_; }; } // namespace ge diff --git a/ge/graph/manager/host_mem_allocator.h b/ge/graph/manager/host_mem_allocator.h index b9dbdc4c..d10b2475 100644 --- a/ge/graph/manager/host_mem_allocator.h +++ b/ge/graph/manager/host_mem_allocator.h @@ -27,7 +27,7 @@ namespace ge { class HostMemAllocator { public: - explicit HostMemAllocator(rtMemType_t) {} + explicit HostMemAllocator(rtMemType_t) {} ~HostMemAllocator() = default; HostMemAllocator(const HostMemAllocator &) = delete; diff --git a/ge/graph/manager/host_mem_manager.cc b/ge/graph/manager/host_mem_manager.cc index c9a33f5c..60a7586d 100644 --- a/ge/graph/manager/host_mem_manager.cc +++ b/ge/graph/manager/host_mem_manager.cc @@ -43,29 +43,20 @@ Status SharedMemAllocator::Allocate(SharedMemInfo &mem_info) { return GE_GRAPH_MEMORY_ALLOC_FAILED; } mem_info.fd = output_para.fd; -#ifndef ONLY_COMPILE_OPEN_SRC mem_info.host_aligned_ptr = AlignedPtr::BuildFromAllocFunc([&output_para](std::unique_ptr &ptr) { ptr.reset(reinterpret_cast(output_para.ptr)); }, [](uint8_t *ptr) { ptr = nullptr; }); -#else - mem_info.host_address = reinterpret_cast(output_para.ptr); -#endif mem_info.device_address = reinterpret_cast(output_para.devPtr); return SUCCESS; } Status SharedMemAllocator::DeAllocate(SharedMemInfo &mem_info) { GELOGD("SharedMemAllocator::DeAllocate"); -#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 - rtFreeHostSharedMemoryIn free_para = {mem_info.shm_name.c_str(), mem_info.mem_size, mem_info.fd, - mem_info.host_address, mem_info.device_address}; -#endif rtError_t rt_ret = rtFreeHostSharedMemory(&free_para); if (rt_ret != RT_ERROR_NONE) { GELOGE(RT_FAILED, "Call rt api(rtFreeHostSharedMemory) failed, ret: 0x%X.", rt_ret); diff --git a/ge/graph/manager/host_mem_manager.h b/ge/graph/manager/host_mem_manager.h index f204c9e4..be3237c3 100644 --- a/ge/graph/manager/host_mem_manager.h +++ b/ge/graph/manager/host_mem_manager.h @@ -42,11 +42,7 @@ struct SharedMemInfo { uint64_t mem_size = 0; int fd = 0; uint8_t *device_address = nullptr; -#ifndef ONLY_COMPILE_OPEN_SRC std::shared_ptr host_aligned_ptr = nullptr; -#else - uint8_t *host_address = nullptr; -#endif SharedMemInfo() = default; SharedMemInfo(string name, uint64_t size) : op_name(std::move(name)), mem_size(size) {} }; diff --git a/ge/graph/optimize/graph_optimize.cc b/ge/graph/optimize/graph_optimize.cc index cd80a956..8cca5b5d 100644 --- a/ge/graph/optimize/graph_optimize.cc +++ b/ge/graph/optimize/graph_optimize.cc @@ -127,6 +127,10 @@ Status GraphOptimize::OptimizeSubGraph(ComputeGraphPtr &compute_graph, const std } Status GraphOptimize::OptimizeOriginalGraph(ComputeGraphPtr &compute_graph) { + if (GetContext().GetHostExecFlag()) { + // graph exec on host, no need OptimizeOriginalGraph + return SUCCESS; + } if (compute_graph == nullptr) { GELOGE(GE_GRAPH_OPTIMIZE_COMPUTE_GRAPH_NULL, "[OptimizeOriginalGraph]: compute_graph is nullptr."); return GE_GRAPH_OPTIMIZE_COMPUTE_GRAPH_NULL; @@ -162,7 +166,7 @@ Status GraphOptimize::OptimizeOriginalGraph(ComputeGraphPtr &compute_graph) { Status GraphOptimize::OptimizeOriginalGraphJudgeInsert(ComputeGraphPtr &compute_graph) { GELOGD("OptimizeOriginalGraphJudgeInsert in"); if (GetContext().GetHostExecFlag()) { - // graph exec on host, no need OptimizeOriginalGraph + // graph exec on host, no need OptimizeOriginalGraphJudgeInsert return SUCCESS; } diff --git a/ge/graph/passes/assign_remove_pass.cc b/ge/graph/passes/assign_remove_pass.cc index 5029b9c3..51e6e006 100644 --- a/ge/graph/passes/assign_remove_pass.cc +++ b/ge/graph/passes/assign_remove_pass.cc @@ -29,7 +29,6 @@ static const std::set kNoTaskNodeTypes = { ge::DATA, ge::ANN_DATA, } namespace ge { -#ifndef ONLY_COMPILE_OPEN_SRC Status AssignRemovePass::Run(NodePtr &node) { GELOGD("AssignRemovePass running"); @@ -145,71 +144,7 @@ Status AssignRemovePass::TransformAttr(NodePtr &node) { } return SUCCESS; } -#else -Status AssignRemovePass::Run(NodePtr &node) { - GELOGD("AssignRemovePass running"); - if (node->GetType() != ASSIGN) { - GELOGD("No need run AssignRemovePass on [%s, %s].", node->GetName().c_str(), node->GetType().c_str()); - return SUCCESS; - } - - const auto &ref_in_anchor = node->GetInDataAnchor(kAssignRefInputIndex); - const auto &value_in_anchor = node->GetInDataAnchor(kAssignValueInputIndex); - if ((ref_in_anchor == nullptr) || (value_in_anchor == nullptr)) { - GELOGE(FAILED, "In data anchor is null, node:%s", node->GetName().c_str()); - return FAILED; - } - const auto &ref_peer_anchor = ref_in_anchor->GetPeerOutAnchor(); - const auto &value_peer_anchor = value_in_anchor->GetPeerOutAnchor(); - if ((ref_peer_anchor == nullptr) || (value_peer_anchor == nullptr)) { - GELOGE(FAILED, "Peer data anchor is null, node:%s", node->GetName().c_str()); - return FAILED; - } - - if (IsCondMatch(node, ref_peer_anchor, value_peer_anchor)) { - /// - /// variable not-const not-const - /// \ / | - /// \ / | - /// Assign ----> variable - /// | | - /// | | - /// node node - /// - GELOGI("Optimization for assign_node %s start", node->GetName().c_str()); - if (IsolateAndDeleteNode(node, {kAssignRefInputIndex}) != SUCCESS) { - GELOGE(FAILED, "Isolate and delete assign_node %s failed.", node->GetName().c_str()); - return FAILED; - } - AddNodeDeleted(node); - - const auto &ref_input = ref_peer_anchor->GetOwnerNode()->GetOpDesc(); - const auto &value_input = value_peer_anchor->GetOwnerNode()->GetOpDesc(); - if ((ref_input == nullptr) || (value_input == nullptr)) { - GELOGE(FAILED, "value input is null"); - return FAILED; - } - if (!AttrUtils::SetStr(value_input->MutableOutputDesc(value_peer_anchor->GetIdx()), ASSIGN_VAR_NAME, - ref_input->GetName())) { - GELOGE(FAILED, "Set attr ASSIGN_VAR_NAME failed."); - return FAILED; - } - // variable has and only has one input - if (ref_input->UpdateInputDesc(0, value_input->GetOutputDesc(value_peer_anchor->GetIdx())) != GRAPH_SUCCESS) { - GELOGE(FAILED, "Update input_desc for variable %s failed.", ref_input->GetName().c_str()); - return FAILED; - } - if (GraphUtils::AddEdge(value_peer_anchor, ref_peer_anchor->GetOwnerNode()->GetInDataAnchor(0)) != GRAPH_SUCCESS) { - GELOGE(FAILED, "Add data edge %s->%s failed", value_input->GetName().c_str(), ref_input->GetName().c_str()); - return FAILED; - } - } - - GELOGD("AssignRemovePass success"); - return SUCCESS; -} -#endif /// /// @brief Check if need optimize for assign_node /// @param [in] assign_node @@ -218,7 +153,7 @@ Status AssignRemovePass::Run(NodePtr &node) { /// @return Status /// bool AssignRemovePass::IsCondMatch(const NodePtr &node, const OutDataAnchorPtr &ref_peer_anchor, - const OutDataAnchorPtr &value_peer_anchor) { + const OutDataAnchorPtr &value_peer_anchor) { GELOGD("Check if assign_node %s match optimization condition, ref_input: %s, value_input: %s", node->GetName().c_str(), ref_peer_anchor->GetOwnerNode()->GetName().c_str(), value_peer_anchor->GetOwnerNode()->GetName().c_str()); diff --git a/ge/graph/passes/assign_remove_pass.h b/ge/graph/passes/assign_remove_pass.h index f8ef2e13..6588df7b 100644 --- a/ge/graph/passes/assign_remove_pass.h +++ b/ge/graph/passes/assign_remove_pass.h @@ -25,7 +25,6 @@ class AssignRemovePass : public BaseNodePass { Status Run(NodePtr &node) override; private: -#ifndef ONLY_COMPILE_OPEN_SRC /// /// @brief Optimize for assign_node /// @param [in] assign_node @@ -39,7 +38,7 @@ class AssignRemovePass : public BaseNodePass { /// @return Status /// Status TransformAttr(NodePtr &node); -#endif + /// /// @brief Check if need optimize for assign_node /// @param [in] assign_node diff --git a/ge/graph/passes/constant_fuse_same_pass.cc b/ge/graph/passes/constant_fuse_same_pass.cc index 8ee89648..eb8b3470 100644 --- a/ge/graph/passes/constant_fuse_same_pass.cc +++ b/ge/graph/passes/constant_fuse_same_pass.cc @@ -115,21 +115,15 @@ void ConstantFuseSamePass::GetFuseConstNodes(ComputeGraphPtr &graph, TypeUtils::DataTypeToSerialString(data_type).c_str()); continue; } -#ifndef ONLY_COMPILE_OPEN_SRC if ((type_size != 0) && (weight->MutableData().GetAlignedPtr() == nullptr)) { GELOGW("aligned_ptr is null while size is not 0"); continue; } -#endif ++insert_const_nums; SameConstKey map_key; map_key.data_size = type_size; -#ifndef ONLY_COMPILE_OPEN_SRC map_key.aligned_ptr = weight->MutableData().GetAlignedPtr(); -#else - map_key.data = weight->GetData().GetData(); -#endif map_key.data_type = data_type; map_key.format = output_tensor->GetFormat(); map_key.shape = output_tensor->GetShape().GetDims(); diff --git a/ge/graph/passes/constant_fuse_same_pass.h b/ge/graph/passes/constant_fuse_same_pass.h index ae39c707..3ff2d6b7 100755 --- a/ge/graph/passes/constant_fuse_same_pass.h +++ b/ge/graph/passes/constant_fuse_same_pass.h @@ -21,20 +21,14 @@ #include #include #include -#ifndef ONLY_COMPILE_OPEN_SRC #include "graph/aligned_ptr.h" -#endif #include "graph/types.h" #include "inc/graph_pass.h" namespace ge { struct SameConstKey { int data_size; -#ifndef ONLY_COMPILE_OPEN_SRC std::shared_ptr aligned_ptr; -#else - const uint8_t *data; -#endif DataType data_type; Format format; std::vector shape; @@ -44,19 +38,12 @@ struct SameConstKey { if (data_size != key.data_size) { return data_size < key.data_size; } -#ifndef ONLY_COMPILE_OPEN_SRC if (data_size != 0) { int ret = memcmp(aligned_ptr->Get(), key.aligned_ptr->Get(), data_size); if (ret != 0) { return ret < 0; } } -#else - int ret = memcmp(data, key.data, data_size); - if (ret != 0) { - return ret < 0; - } -#endif if (data_type != key.data_type) { return data_type < key.data_type; } diff --git a/ge/graph/preprocess/graph_preprocess.cc b/ge/graph/preprocess/graph_preprocess.cc index f94633a1..0bfec241 100644 --- a/ge/graph/preprocess/graph_preprocess.cc +++ b/ge/graph/preprocess/graph_preprocess.cc @@ -38,9 +38,6 @@ #include "graph/passes/aicpu_constant_folding_pass.h" #include "graph/passes/assert_pass.h" #include "ge/ge_api_types.h" -#ifdef ONLY_COMPILE_OPEN_SRC -#include "graph/passes/assign_remove_pass.h" -#endif #include "graph/passes/common_subexpression_elimination_pass.h" #include "graph/passes/cond_pass.h" #include "graph/passes/cond_remove_pass.h" @@ -1865,9 +1862,6 @@ Status GraphPrepare::PrepareOptimize() { VarIsInitializedOpPass var_is_initialized_pass; ParallelConcatStartOpPass parallel_concat_start_op_pass; IdentityPass identity_pass(false); -#ifdef ONLY_COMPILE_OPEN_SRC - AssignRemovePass assign_remove_pass; -#endif SnapshotPass snapshot_pass; if (!options_.train_graph_flag) { names_to_passes.emplace_back("DropOutPass", &dropout_pass); @@ -1882,11 +1876,6 @@ Status GraphPrepare::PrepareOptimize() { names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass); names_to_passes.emplace_back("ParallelConcatStartOpPass", ¶llel_concat_start_op_pass); names_to_passes.emplace_back("IdentityPass", &identity_pass); -#ifdef ONLY_COMPILE_OPEN_SRC - if (GetContext().GetHostExecFlag()) { - names_to_passes.emplace_back("AssignRemovePass", &assign_remove_pass); - } -#endif GE_TIMESTAMP_START(names_to_passes); ret = ge_passes.Run(names_to_passes); GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses"); diff --git a/ge/hybrid/common/npu_memory_allocator.cc b/ge/hybrid/common/npu_memory_allocator.cc index c2602f37..ccd6a624 100644 --- a/ge/hybrid/common/npu_memory_allocator.cc +++ b/ge/hybrid/common/npu_memory_allocator.cc @@ -20,9 +20,7 @@ #include "graph/manager/graph_caching_allocator.h" #include "graph/manager/graph_mem_allocator.h" #include "graph/manager/rdma_pool_allocator.h" -#ifndef ONLY_COMPILE_OPEN_SRC #include "graph/manager/host_mem_allocator.h" -#endif namespace ge { namespace hybrid { @@ -67,11 +65,7 @@ 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) { -#ifndef ONLY_COMPILE_OPEN_SRC buffer = MemManager::Instance().HostMemInstance(RT_MEMORY_HBM).Malloc(allocate_size); -#else - buffer = malloc(allocate_size); -#endif } else { if (allocate_size > kMaxHbmMemorySize) { GELOGE(PARAM_INVALID, "Invalid HBM memory size: %zu", allocate_size); @@ -108,11 +102,7 @@ void NpuMemoryAllocator::Deallocate(void *data, MemStorageType mem_type) { if (mem_type == RDMA_HBM) { MemManager::Instance().RdmaPoolInstance(RT_MEMORY_HBM).Free(reinterpret_cast(data), device_id_); } else if (mem_type == HOST_DDR) { -#ifndef ONLY_COMPILE_OPEN_SRC MemManager::Instance().HostMemInstance(RT_MEMORY_HBM).Free(data); -#else - free(data); -#endif } else { MemManager::Instance().CachingInstance(RT_MEMORY_HBM).Free(reinterpret_cast(data), device_id_); } diff --git a/ge/hybrid/model/hybrid_model_builder.cc b/ge/hybrid/model/hybrid_model_builder.cc index 32fc495a..d1f61985 100755 --- a/ge/hybrid/model/hybrid_model_builder.cc +++ b/ge/hybrid/model/hybrid_model_builder.cc @@ -25,10 +25,8 @@ #include "graph/manager/graph_var_manager.h" #include "graph/manager/host_mem_manager.h" #include "graph/manager/trans_var_data_utils.h" -#ifndef ONLY_COMPILE_OPEN_SRC #include "graph/manager/graph_mem_allocator.h" #include "graph/manager/host_mem_allocator.h" -#endif #include "graph/utils/graph_utils.h" #include "hybrid/common/npu_memory_allocator.h" #include "hybrid/node_executor/node_executor.h" @@ -865,7 +863,6 @@ Status HybridModelBuilder::InitConstantOps() { std::unique_ptr var_tensor; if (GetContext().GetHostExecFlag()) { -#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 @@ -878,11 +875,6 @@ Status HybridModelBuilder::InitConstantOps() { } var_tensor.reset(new(std::nothrow)TensorValue(aligned_tensor.MutableData().data(), aligned_tensor.GetData().size())); -#else - auto buffer = ge_tensor->MutableData(); - GELOGD("Init tensor with host constant. size = %zu", buffer.GetSize()); - var_tensor.reset(new(std::nothrow)TensorValue(buffer.GetData(), buffer.GetSize())); -#endif } else { GE_CHK_STATUS_RET_NOLOG(VarNodeToTensor(var_node, var_tensor)); GELOGD("Init const op tensor. name = %s, size = %ld", var_name.c_str(), var_tensor->GetSize()); @@ -937,7 +929,6 @@ Status HybridModelBuilder::InitVariableTensors() { GELOGE(GE_GRAPH_MALLOC_FAILED, "Host variable [%s] malloc failed.", it.first.c_str()); return GE_GRAPH_MALLOC_FAILED; } -#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."); @@ -947,11 +938,6 @@ Status HybridModelBuilder::InitVariableTensors() { std::unique_ptr tensor(new (std::nothrow) TensorValue(mem_info.host_aligned_ptr->MutableGet(), tensor_size)); -#else - GELOGD("Host variable [%s] malloc success.", it.first.c_str()); - - std::unique_ptr tensor(new (std::nothrow) TensorValue(mem_info.host_address, tensor_size)); -#endif GE_CHECK_NOTNULL(tensor); hybrid_model_.variable_tensors_.emplace(it.first, std::move(tensor)); } diff --git a/ge/hybrid/node_executor/host_cpu/host_cpu_node_executor.cc b/ge/hybrid/node_executor/host_cpu/host_cpu_node_executor.cc index 32522fe8..0cc635e4 100755 --- a/ge/hybrid/node_executor/host_cpu/host_cpu_node_executor.cc +++ b/ge/hybrid/node_executor/host_cpu/host_cpu_node_executor.cc @@ -18,10 +18,8 @@ #include "hybrid/node_executor/host_cpu/kernel_factory.h" #include "graph/passes/folding_pass.h" #include "hybrid/model/hybrid_model.h" -#ifndef ONLY_COMPILE_OPEN_SRC #include "graph/manager/graph_mem_allocator.h" #include "graph/manager/host_mem_allocator.h" -#endif #include "ge_local_engine/engine/host_cpu_engine.h" namespace ge { @@ -54,18 +52,11 @@ 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; -#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()); GE_CHECK_NOTNULL(item.second); auto in_tensor = MakeShared(input_desc, item.second, item.first); -#else - GE_CHECK_NOTNULL(context.GetInput(i)); - auto in_tensor = MakeShared(input_desc, - reinterpret_cast(context.GetInput(i)->GetData()), - context.GetInput(i)->GetSize()); -#endif GE_CHECK_NOTNULL(in_tensor); in_tensor->MutableTensorDesc().SetDataType(input_desc.GetDataType()); in_tensor->MutableTensorDesc().SetShape(input_desc.GetShape()); @@ -84,15 +75,9 @@ Status CpuKernelNodeTask::Execute(TaskContext &context) { } auto tensor = context.GetOutput(i); GE_CHECK_NOTNULL(tensor); -#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(output_desc, item.second, item.first); -#else - auto out_tensor = MakeShared(output_desc, - reinterpret_cast(tensor->GetData()), - tensor->GetSize()); -#endif GE_CHECK_NOTNULL(out_tensor); out_tensor->MutableTensorDesc().SetDataType(output_desc.GetDataType()); out_tensor->MutableTensorDesc().SetShape(output_desc.GetShape());