diff --git a/ge/ge_local_engine/engine/host_cpu_engine.cc b/ge/ge_local_engine/engine/host_cpu_engine.cc index cd68ae15..d9b67736 100755 --- a/ge/ge_local_engine/engine/host_cpu_engine.cc +++ b/ge/ge_local_engine/engine/host_cpu_engine.cc @@ -13,15 +13,15 @@ * See the License for the specific language governing permissions and * limitations under the License. */ -#include "host_cpu_engine.h" -#include "graph/common/omg_util.h" +#include "ge_local_engine/engine/host_cpu_engine.h" #include "graph/utils/op_desc_utils.h" #include "graph/utils/tensor_adapter.h" +#include "graph/utils/node_utils.h" +#include "graph/utils/type_utils.h" #include "register/op_kernel_registry.h" #include "register/host_cpu_context.h" #include "common/ge/ge_util.h" #include "common/ge/plugin_manager.h" -#include "graph/utils/type_utils.h" #include "common/fp16_t.h" #include "common/math/math_util.h" @@ -123,10 +123,7 @@ bool HostCpuEngine::CheckSupported(const string &op_type) { } Status HostCpuEngine::FindOpKernel(const ge::NodePtr &node, std::unique_ptr &op_kernel) { - std::string op_type; - auto status = GetOriginalType(node, op_type); - GE_CHK_BOOL_EXEC_NOLOG(status == SUCCESS, return status); - + const std::string op_type = NodeUtils::GetNodeType(node); auto kernel = OpKernelRegistry::GetInstance().CreateHostCpuOp(op_type); if (kernel == nullptr) { GELOGD("Op of type %s is not supported by host cpu engine", op_type.c_str()); diff --git a/ge/graph/load/model_manager/model_manager.cc b/ge/graph/load/model_manager/model_manager.cc index cae828d6..6fd2e273 100755 --- a/ge/graph/load/model_manager/model_manager.cc +++ b/ge/graph/load/model_manager/model_manager.cc @@ -1378,7 +1378,9 @@ Status ModelManager::LoadCustAicpuSo(const OpDescPtr &op_desc, const string &so_ Status ModelManager::LaunchKernelCustAicpuSo(const string &kernel_name) { GELOGD("Aicpu kernel launch task in, kernel name %s.", kernel_name.c_str()); std::lock_guard lock(cust_aicpu_mutex_); - if (cust_aicpu_so_.size() == 0) return SUCCESS; + if (cust_aicpu_so_.empty()) { + return SUCCESS; + } // get current context rtContext_t rt_cur_ctx = nullptr; auto rt_error = rtCtxGetCurrent(&rt_cur_ctx); @@ -1394,17 +1396,19 @@ Status ModelManager::LaunchKernelCustAicpuSo(const string &kernel_name) { return SUCCESS; } - vector allocated_mem; - rtError_t status; rtStream_t stream = nullptr; + vector allocated_mem; std::function callback = [&]() { for (auto mem : allocated_mem) { GE_CHK_RT(rtFree(mem)); } - GE_CHK_RT(rtStreamDestroy(stream)); + if (stream != nullptr) { + GE_CHK_RT(rtStreamDestroy(stream)); + } }; GE_MAKE_GUARD(release, callback); + rtError_t status; vector v_cust_so; void *args = nullptr; diff --git a/ge/graph/partition/dynamic_shape_partition.cc b/ge/graph/partition/dynamic_shape_partition.cc index 1db47498..8fc19ff2 100755 --- a/ge/graph/partition/dynamic_shape_partition.cc +++ b/ge/graph/partition/dynamic_shape_partition.cc @@ -284,9 +284,6 @@ Status DynamicShapePartitioner::InitClusters() { auto cluster = MakeShared(rank++, type, node, this); REQUIRE_NOT_NULL(cluster, "[New][Memory] for cluster failed."); node_2_cluster_[node] = cluster; - if (cluster->IsUnknownShape()) { - ordered_cluster_.push_back(cluster); - } int64_t group_index = -1; if (AttrUtils::GetInt(node->GetOpDesc(), ATTR_NAME_CONTROL_FLOW_GROUP, group_index)) { @@ -306,7 +303,7 @@ Status DynamicShapePartitioner::InitClusters() { return SUCCESS; } -Status DynamicShapePartitioner::TopologicalSortClusters() { +Status DynamicShapePartitioner::TopologicalSortClusters(const OrderedFilter &ordered_filter) { ordered_cluster_.clear(); // BFS topological sort clusters for known shape cluster std::queue ready_clusters; @@ -331,7 +328,7 @@ Status DynamicShapePartitioner::TopologicalSortClusters() { auto cluster = ready_clusters.front(); ready_clusters.pop(); cluster->UpdateRank(rank++); - if (cluster->IsKnownShape() || cluster->IsInputNode()) { + if (ordered_filter == nullptr || ordered_filter(cluster)) { ordered_cluster_.push_back(cluster); } for (const auto &out_cluster : cluster->Outputs()) { @@ -378,7 +375,6 @@ void DynamicShapePartitioner::MergeClustersControlFlow() { continue; } - bool is_unknown_cluster = cluster->IsUnknownShape(); for (++rit; rit != control_cluster.rend(); ++rit) { const auto &cluster_from = *rit; if (all_merged_clusters.count(cluster_from) > 0) { @@ -395,11 +391,6 @@ void DynamicShapePartitioner::MergeClustersControlFlow() { } } } - - if (!is_unknown_cluster && cluster->IsUnknownShape()) { - GELOGD("Add to ordered cluster: %s", cluster->DebugString().c_str()); - ordered_cluster_.push_back(cluster); - } } } @@ -475,9 +466,19 @@ void DynamicShapePartitioner::MergeClustersInputData() { } Status DynamicShapePartitioner::MergeClusters() { + const auto filter_known = [](const ClusterPtr &cluster) { + return cluster->IsKnownShape() || cluster->IsInputNode(); + }; + const auto filter_unknown = [](const ClusterPtr &cluster) { + return cluster->IsUnknownShape(); + }; + MergeClustersControlFlow(); + REQUIRE_SUCCESS(TopologicalSortClusters(filter_unknown), + "[TopologicalSort][Clusters] after merge control flow clusters failed."); MergeClustersUnknownShape(); - REQUIRE_SUCCESS(TopologicalSortClusters(), "[TopologicalSort][Clusters] after merge unknown shape clusters failed."); + REQUIRE_SUCCESS(TopologicalSortClusters(filter_known), + "[TopologicalSort][Clusters] after merge unknown shape clusters failed."); MergeClustersKnownShape(); MergeClustersInputData(); return SUCCESS; diff --git a/ge/graph/partition/dynamic_shape_partition.h b/ge/graph/partition/dynamic_shape_partition.h index bd3b128f..7e82d131 100644 --- a/ge/graph/partition/dynamic_shape_partition.h +++ b/ge/graph/partition/dynamic_shape_partition.h @@ -111,6 +111,8 @@ class DynamicShapePartitioner { Status Partition(); + using OrderedFilter = std::function &cluster)>; + private: Status PartitionImpl(); // Collect nodes that satisfy the unknowshape rules: @@ -138,7 +140,7 @@ class DynamicShapePartitioner { // Merge clusters step3 void MergeClustersInputData(); // Topological sort clusters after merge unknown shape clusters. - Status TopologicalSortClusters(); + Status TopologicalSortClusters(const OrderedFilter &ordered_filter); // Deduplicate merged clusters void PruneUniqueClusters(); // Establish the input-output anchors for each partition of the cluster and record links to other clusters diff --git a/ge/graph/passes/mark_force_unknown_for_cond_pass.cc b/ge/graph/passes/mark_force_unknown_for_cond_pass.cc index 74babadc..e024217f 100644 --- a/ge/graph/passes/mark_force_unknown_for_cond_pass.cc +++ b/ge/graph/passes/mark_force_unknown_for_cond_pass.cc @@ -16,8 +16,6 @@ #include "mark_force_unknown_for_cond_pass.h" -#include - #include "graph/utils/node_utils.h" #include "graph/common/omg_util.h" @@ -26,17 +24,7 @@ namespace { inline bool IsMergeInLoop(const NodePtr &node) { const static std::set kLoopMergeInputs{ ENTER, REFENTER, NEXTITERATION, REFNEXTITERATION }; - std::string node_type; - (void)GetOriginalType(node, node_type); - return kLoopMergeInputs.count(node_type) > 0; -} - -inline bool IsSwitchInLoop(const NodePtr &node) { - const static std::set kLoopSwitchInputs{ MERGE, REFMERGE, LOOPCOND }; - - std::string node_type; - (void)GetOriginalType(node, node_type); - return kLoopSwitchInputs.count(node_type) > 0; + return kLoopMergeInputs.count(NodeUtils::GetNodeType(node)) > 0; } } @@ -44,10 +32,7 @@ Status MarkForceUnknownForCondPass::Run(ComputeGraphPtr graph) { GELOGD("MarkForceUnknownForCondPass Enter"); std::map> switch_groups; for (const auto &node : graph->GetDirectNode()) { - std::string node_type; - GE_CHK_STATUS_RET(GetOriginalType(node, node_type), - "[Get][OriginalType] of node in graph:%s failed.", graph->GetName().c_str()); - if (kMergeOpTypes.count(node_type) == 0) { + if (kMergeOpTypes.count(NodeUtils::GetNodeType(node)) == 0) { continue; } @@ -64,6 +49,51 @@ Status MarkForceUnknownForCondPass::Run(ComputeGraphPtr graph) { return SUCCESS; } +/// +/// @brief Deal with Switch node for LoopCond +/// @param [in] Switch node +/// @param [in] dest span +/// @param [out] Search queue +/// @return true: Switch In while loop / false: Not in while Loop. +/// +bool MarkForceUnknownForCondPass::DealAsLoopSwitch(const NodePtr &node, uint32_t dst_span, + std::queue> &search_queue) { + /// LoopCond --->\. + /// \. + /// Enter-----------+ \. + /// +--> Merge --> Switch --> Exit + /// NextIteration---+ + const auto is_loop_op = [](const NodePtr &n) { + return NodeUtils::GetNodeType(n) == LOOPCOND; + }; + const auto is_exit_op = [](const NodePtr &n) { + return kExitOpTypes.count(NodeUtils::GetNodeType(n)) > 0; + }; + + const auto src_nodes = node->GetInAllNodes(); + const auto dst_nodes = node->GetOutAllNodes(); + if (std::none_of(src_nodes.begin(), src_nodes.end(), is_loop_op) && + std::none_of(dst_nodes.begin(), dst_nodes.end(), is_exit_op)) { + return false; + } + + for (const auto &m : src_nodes) { + if (kMergeOpTypes.count(NodeUtils::GetNodeType(m)) > 0) { + for (const auto &n : m->GetInAllNodes()) { + if (kNextIterationOpTypes.count(NodeUtils::GetNodeType(n)) > 0) { + continue; + } + + search_queue.push({n, dst_span}); + GELOGD("Travel in Loop: %s <-- %s <-- %s, span is: %u", node->GetName().c_str(), m->GetName().c_str(), + n->GetName().c_str(), dst_span); + } + } + } + + return true; +} + /// /// @brief Mark force unknown shape for Switch node /// @param [in] merge node @@ -72,6 +102,7 @@ Status MarkForceUnknownForCondPass::Run(ComputeGraphPtr graph) { /// void MarkForceUnknownForCondPass::MarkUnknownForSwitch(const NodePtr &node, std::vector &switch_group) { // Switch --> {Switch --> Merge} --> Merge + GELOGD("Search Switch node for Merge: %s", node->GetName().c_str()); std::unordered_set nodes_seen; std::queue> search_queue({{node, 0}}); while (!search_queue.empty()) { @@ -79,43 +110,25 @@ void MarkForceUnknownForCondPass::MarkUnknownForSwitch(const NodePtr &node, std: const auto dst_span = search_queue.front().second; search_queue.pop(); - // Switch --> Identity --> Constant - for (const auto &in_node : dst_node->GetInControlNodes()) { - if (nodes_seen.count(in_node) > 0) { - GELOGD("Travel node: %s, Skip already seen node: %s", dst_node->GetName().c_str(), in_node->GetName().c_str()); - continue; - } - nodes_seen.insert(in_node); - - if (in_node->GetType() == IDENTITY) { - GELOGD("Travel node: %s, In control: %s, span is: %u", dst_node->GetName().c_str(), - in_node->GetName().c_str(), dst_span); - search_queue.push({in_node, dst_span}); - } - } - - for (const auto &in_node : dst_node->GetInDataNodes()) { + for (const auto &in_node : dst_node->GetInAllNodes()) { if (nodes_seen.count(in_node) > 0) { GELOGD("Travel node: %s, Skip already seen node: %s", dst_node->GetName().c_str(), in_node->GetName().c_str()); continue; } nodes_seen.insert(in_node); - std::string node_type; - (void)GetOriginalType(in_node, node_type); + const std::string node_type = NodeUtils::GetNodeType(in_node); GELOGD("Travel node: %s, %s node: %s, span is: %u", dst_node->GetName().c_str(), node_type.c_str(), in_node->GetName().c_str(), dst_span); if (kSwitchOpTypes.count(node_type) > 0) { // Switch input node. + if (DealAsLoopSwitch(in_node, dst_span, search_queue)) { + continue; + } + if (dst_span > 0) { search_queue.push({in_node, dst_span - 1}); } else { - const auto &all_in_nodes = in_node->GetInDataNodes(); - if (std::any_of(all_in_nodes.begin(), all_in_nodes.end(), IsSwitchInLoop)) { - GELOGW("Travel node: %s, %s node: %s, Skip LoopCond switch", dst_node->GetName().c_str(), node_type.c_str(), - in_node->GetName().c_str()); - } else { - switch_group.emplace_back(in_node); - } + switch_group.emplace_back(in_node); } } else if (kMergeOpTypes.count(node_type) > 0) { // Merge input node. search_queue.push({in_node, dst_span + 1}); diff --git a/ge/graph/passes/mark_force_unknown_for_cond_pass.h b/ge/graph/passes/mark_force_unknown_for_cond_pass.h index 528a8fdc..030b55ee 100644 --- a/ge/graph/passes/mark_force_unknown_for_cond_pass.h +++ b/ge/graph/passes/mark_force_unknown_for_cond_pass.h @@ -19,12 +19,23 @@ #include "inc/graph_pass.h" +#include + namespace ge { class MarkForceUnknownForCondPass : public GraphPass { public: Status Run(ComputeGraphPtr graph); private: + /// + /// @brief Deal with Switch node for LoopCond + /// @param [in] Switch node + /// @param [in] dest span + /// @param [out] Search queue + /// @return true: Switch In while loop / false: Not in while Loop. + /// + bool DealAsLoopSwitch(const NodePtr &node, uint32_t dst_span, std::queue> &search_queue); + /// /// @brief Mark force unknown shape for Switch node /// @param [in] merge node diff --git a/ge/graph/passes/next_iteration_pass.cc b/ge/graph/passes/next_iteration_pass.cc index fb8f8627..1c2d7218 100644 --- a/ge/graph/passes/next_iteration_pass.cc +++ b/ge/graph/passes/next_iteration_pass.cc @@ -24,7 +24,9 @@ using std::string; namespace ge { namespace { -const int64_t kLoopType = 1; +constexpr int64_t kLoopType = 1; +constexpr uint8_t kMaxTransOp = 3; +constexpr uint8_t kTransOpIoSize = 1; } Status NextIterationPass::Run(ComputeGraphPtr graph) { @@ -287,18 +289,25 @@ void NextIterationPass::HandleSwitchExitNodes(const LoopCondGroup &loop_group, i std::string node_type; for (const auto &switch_node : loop_group.switch_nodes) { SetControlFlowGroup(switch_node, group_index); - for (const auto &node : switch_node->GetOutDataNodes()) { - (void)GetOriginalType(node, node_type); - if (kExitOpTypes.count(node_type) > 0) { - SetControlFlowGroup(node, group_index); - } else { - // For: Switch -> Cast -> Exit - for (const auto &n : node->GetOutDataNodes()) { - (void)GetOriginalType(n, node_type); - if (kExitOpTypes.count(node_type) > 0) { - SetControlFlowGroup(n, group_index); - } + for (auto node : switch_node->GetOutDataNodes()) { + // Switch --> Exit + // Switch --> Cast --> Exit + // Switch --> TransData --> Cast --> Exit + for (uint8_t i = 0; i < kMaxTransOp; ++i) { + if (node->GetInDataNodes().size() != kTransOpIoSize || node->GetAllOutDataAnchorsSize() != kTransOpIoSize) { + break; } + + if (kExitOpTypes.count(NodeUtils::GetNodeType(node)) > 0) { + SetControlFlowGroup(node, group_index); + break; + } + + const auto &all_nodes = node->GetOutAllNodes(); + if (all_nodes.size() != kTransOpIoSize) { + break; + } + node = all_nodes.at(0); } } } diff --git a/ge/graph/passes/switch_to_stream_switch_pass.cc b/ge/graph/passes/switch_to_stream_switch_pass.cc index e4ab0111..1a47c14b 100644 --- a/ge/graph/passes/switch_to_stream_switch_pass.cc +++ b/ge/graph/passes/switch_to_stream_switch_pass.cc @@ -395,8 +395,9 @@ NodePtr SwitchToStreamSwitchPass::CreateStreamSwitchNode(const ComputeGraphPtr & peer_cond_anchor->GetOwnerNode()->GetName().c_str(), stream_switch->GetName().c_str()); int64_t group_index = -1; - (void)AttrUtils::GetInt(switch_node->GetOpDesc(), ATTR_NAME_CONTROL_FLOW_GROUP, group_index); - SetControlFlowGroup(stream_switch, group_index); + if (AttrUtils::GetInt(switch_node->GetOpDesc(), ATTR_NAME_CONTROL_FLOW_GROUP, group_index)) { + SetControlFlowGroup(stream_switch, group_index); + } return stream_switch; } diff --git a/ge/hybrid/executor/node_state.cc b/ge/hybrid/executor/node_state.cc index 468c84e6..ad38c792 100644 --- a/ge/hybrid/executor/node_state.cc +++ b/ge/hybrid/executor/node_state.cc @@ -326,17 +326,45 @@ std::shared_ptr NodeState::GetTaskContext() { } void NodeState::SavePersistTensor(int input_idx, const TensorValue &tensor) { - if (node_item_->root_data_.count(input_idx) > 0) { - GELOGD("[%s] Save Root input tensor: %d", GetName().c_str(), input_idx); - root_tensor_values_[input_idx] = tensor; + const auto is_persist_tensor = [](const std::map> &items, int idx) { + const auto is_exist = [&idx](const std::pair> &items) { + return items.second.count(idx) > 0; + }; + return std::any_of(items.begin(), items.end(), is_exist); + }; + + if (root_tensor_values_.count(input_idx) > 0) { + return; } - if (node_item_->enter_data_.count(input_idx) > 0) { + if (is_persist_tensor(node_item_->root_data_, input_idx)) { + GELOGD("[%s] Save Root input tensor: %d", GetName().c_str(), input_idx); + root_tensor_values_[input_idx] = tensor; + } else if (is_persist_tensor(node_item_->enter_data_, input_idx)) { GELOGD("[%s] Save Enter input tensor: %d", GetName().c_str(), input_idx); root_tensor_values_[input_idx] = tensor; } } +void NodeState::UpdatePersistTensor() { + const auto update_tensor = [&](const std::map> &items) { + for (const auto &item : items) { + for (const auto idx : item.second) { + UpdatePersistTensor(idx); + } + } + }; + + if (root_tensor_values_.empty()) { + return; + } + + update_tensor(node_item_->root_data_); + if (iteration_count_ > 0) { + update_tensor(node_item_->enter_data_); + } +} + void NodeState::UpdatePersistTensor(int input_idx) { const auto it = root_tensor_values_.find(input_idx); if (it == root_tensor_values_.end()) { @@ -363,16 +391,9 @@ void NodeState::ResetContext(uint64_t iteration) { data_scheduled_ = static_cast(node_item_->root_data_.size()); ctrl_scheduled_ = static_cast(node_item_->root_ctrl_.size()); - for (auto item : node_item_->root_data_) { - UpdatePersistTensor(item.first); - } - if (iteration > 0) { data_scheduled_ += static_cast(node_item_->enter_data_.size()); ctrl_scheduled_ += static_cast(node_item_->enter_ctrl_.size()); - for (auto item : node_item_->enter_data_) { - UpdatePersistTensor(item.first); - } } iteration_count_ = iteration; diff --git a/ge/hybrid/executor/node_state.h b/ge/hybrid/executor/node_state.h index b80b60b0..1ec8517e 100644 --- a/ge/hybrid/executor/node_state.h +++ b/ge/hybrid/executor/node_state.h @@ -132,6 +132,7 @@ struct NodeState { void RunNextIteration(); void SavePersistTensor(int input_idx, const TensorValue &tensor); + void UpdatePersistTensor(); Status NodeScheduled(const std::function &ready) const; diff --git a/ge/hybrid/executor/worker/execution_engine.cc b/ge/hybrid/executor/worker/execution_engine.cc index 8eecbc80..ca864244 100755 --- a/ge/hybrid/executor/worker/execution_engine.cc +++ b/ge/hybrid/executor/worker/execution_engine.cc @@ -373,6 +373,7 @@ Status ExecutionEngine::DoExecuteAsync(NodeState &node_state, auto executor = node_item.node_executor; GE_CHECK_NOTNULL(executor); RECORD_EXECUTION_EVENT(&context, task_context.GetNodeName(), "[PrepareTask] Start"); + node_state.UpdatePersistTensor(); GE_CHK_STATUS_RET(executor->PrepareTask(*task, task_context), "[Prepare][Task] for [%s] failed.", node_state.GetName().c_str()); RECORD_EXECUTION_EVENT(&context, task_context.GetNodeName(), "[PrepareTask] End"); diff --git a/ge/hybrid/model/hybrid_model_builder.cc b/ge/hybrid/model/hybrid_model_builder.cc index 6a966622..4abf20c8 100755 --- a/ge/hybrid/model/hybrid_model_builder.cc +++ b/ge/hybrid/model/hybrid_model_builder.cc @@ -288,10 +288,6 @@ Status HybridModelBuilder::GetOrCreateNodeItem(const NodePtr &node, NodeItem **n return SUCCESS; } - if (node->GetType() == MEMCPYASYNC) { // Convert MemcpyAsync to Identity. - node->GetOpDesc()->SetType(IDENTITY); - } - std::unique_ptr new_node; GE_CHK_STATUS_RET(NodeItem::Create(node, new_node), "[Invoke][Create] failed, model_name_:[%s]", GetGraphName()); GE_CHK_STATUS_RET_NOLOG(NodeExecutorManager::GetInstance().GetExecutor(*node, &new_node->node_executor)); diff --git a/ge/hybrid/model/node_item.cc b/ge/hybrid/model/node_item.cc index cef06fc6..fb7add48 100644 --- a/ge/hybrid/model/node_item.cc +++ b/ge/hybrid/model/node_item.cc @@ -14,10 +14,8 @@ * limitations under the License. */ -#include "node_item.h" -#include -#include "common/debug/log.h" -#include "graph/common/omg_util.h" +#include "hybrid/model/node_item.h" + #include "graph/compute_graph.h" #include "graph/debug/ge_attr_define.h" #include "hybrid/executor/worker/shape_inference_engine.h" @@ -26,6 +24,8 @@ namespace ge { namespace hybrid { namespace { +const uint8_t kMaxTransCount = 3; +const uint32_t kTransOpIoSize = 1; const char *const kAttrNameOriginalFusionGraph = "_original_fusion_graph"; const char *const kNodeTypeRetVal = "_RetVal"; const std::set kControlOpTypes{ @@ -41,6 +41,25 @@ const std::set kMergeOpTypes{ MERGE, REFMERGE, STREAMMERGE }; +bool IsEnterFeedNode(NodePtr node) { + // For: Enter -> node + // For: Enter -> Cast -> node + // For: Enter -> TransData -> Cast -> node + for (uint8_t i = 0; i < kMaxTransCount; ++i) { + if (kEnterOpTypes.count(NodeUtils::GetNodeType(node)) > 0) { + GELOGD("Node[%s] is Enter feed node.", node->GetName().c_str()); + return true; + } + + const auto all_nodes = node->GetInDataNodes(); + if (all_nodes.size() != kTransOpIoSize || node->GetAllInDataAnchorsSize() != kTransOpIoSize) { + return false; + } + node = all_nodes.at(0); + } + return false; +} + Status ParseInputMapping(Node &node, OpDesc &op_desc, FusedSubgraph &fused_subgraph) { uint32_t parent_index = 0; if (!AttrUtils::GetInt(op_desc, ATTR_NAME_PARENT_NODE_INDEX, parent_index)) { @@ -98,8 +117,7 @@ Status ParseFusedSubgraph(NodeItem &node_item) { GE_CHECK_NOTNULL(node); auto op_desc = node->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); - std::string node_type; - GE_CHK_STATUS_RET(GetOriginalType(node, node_type)); + const std::string node_type = NodeUtils::GetNodeType(node); if (node_type == DATA) { GE_CHK_GRAPH_STATUS_RET(ParseInputMapping(*node, *op_desc, *fused_subgraph)); } else if (node_type == kNodeTypeRetVal) { @@ -398,19 +416,21 @@ void NodeItem::SetDataSend(NodeItem *node_item, int anchor_index) { data_send_.emplace(node_item); node_item->data_recv_[this] = anchor_index; if (is_root_node_) { - node_item->root_data_[anchor_index] = this; + auto &data_anchors = node_item->root_data_[this]; + data_anchors.emplace(anchor_index); } // If Enter feed Not Merge, take as root Node. - if (IsEnterOp() && (node_item->node_type != STREAMMERGE)) { - node_item->enter_data_[anchor_index] = this; + if (IsEnterFeedNode(node) && (node_item->node_type != STREAMMERGE)) { + auto &data_anchors = node_item->enter_data_[this]; + data_anchors.emplace(anchor_index); } GELOGI("Node[%s] will control node[%s]", NodeName().c_str(), node_item->NodeName().c_str()); } void NodeItem::SetCtrlSend(NodeItem *node_item, uint32_t switch_index) { if (switch_index < switch_groups_.size()) { - std::vector &switch_group = switch_groups_[switch_index]; - switch_group.emplace_back(node_item); + auto &switch_group = switch_groups_[switch_index]; + switch_group.emplace(node_item); } else { ctrl_send_.insert(node_item); } @@ -420,7 +440,7 @@ void NodeItem::SetCtrlSend(NodeItem *node_item, uint32_t switch_index) { node_item->root_ctrl_.emplace(this); } // If Enter feed control signal, take as root Node. - if (IsEnterOp() && (node_item->node_type != STREAMMERGE && node_item->node_type != STREAMACTIVE)) { + if (IsEnterFeedNode(node) && (node_item->node_type != STREAMMERGE && node_item->node_type != STREAMACTIVE)) { node_item->enter_ctrl_.emplace(this); } GELOGI("Node[%s] will control node[%s]", NodeName().c_str(), node_item->NodeName().c_str()); @@ -433,8 +453,8 @@ void NodeItem::SetMergeCtrl(NodeItem *node_item, uint32_t merge_index) { } // this is StreamMerge node, node_item is StreamActive node. - std::vector &switch_group = switch_groups_[merge_index]; - switch_group.emplace_back(node_item); + auto &switch_group = switch_groups_[merge_index]; + switch_group.emplace(node_item); node_item->ctrl_send_.emplace(this); GELOGI("Node[%s] will control node[%s]", node_item->NodeName().c_str(), NodeName().c_str()); diff --git a/ge/hybrid/model/node_item.h b/ge/hybrid/model/node_item.h index ec66f094..f6dcdcf6 100644 --- a/ge/hybrid/model/node_item.h +++ b/ge/hybrid/model/node_item.h @@ -148,14 +148,14 @@ struct NodeItem { int64_t frame_index_ = -1; int64_t parent_frame_ = -1; std::set root_ctrl_; // Recv ctrl from root node - std::map root_data_; // Recv data from root node + std::map> root_data_; // Recv data from root node std::set enter_ctrl_; // Recv ctrl from Enter node - std::map enter_data_; // Recv data from Enter node + std::map> enter_data_; // Recv data from Enter node std::set data_send_; // Send data notify to std::map data_recv_; // Recv data notify from std::set ctrl_send_; // Send ctrl notify to std::set ctrl_recv_; // Recv ctrl notify from - std::vector> switch_groups_; // Send ctrl notify to + std::vector> switch_groups_; // Send ctrl notify to std::shared_ptr kernel_task; std::unique_ptr fused_subgraph; diff --git a/ge/hybrid/node_executor/hccl/hccl_node_executor.cc b/ge/hybrid/node_executor/hccl/hccl_node_executor.cc index d942695e..0f98674c 100644 --- a/ge/hybrid/node_executor/hccl/hccl_node_executor.cc +++ b/ge/hybrid/node_executor/hccl/hccl_node_executor.cc @@ -342,6 +342,7 @@ Status RdmaNodeTask::ExecuteAsync(TaskContext &context, std::function do GE_CHK_RT_RET(rtEventDestroy(evt)); } GELOGI("rdma callback success."); + return SUCCESS; }; HcclResult hccl_ret = HcomExecEnqueueRemoteAccess(context.GetNodeItem().NodeType(), addr_infos, callback); diff --git a/ge/hybrid/node_executor/rts/rts_node_executor.cc b/ge/hybrid/node_executor/rts/rts_node_executor.cc index 3ad791b6..e3058ee3 100644 --- a/ge/hybrid/node_executor/rts/rts_node_executor.cc +++ b/ge/hybrid/node_executor/rts/rts_node_executor.cc @@ -17,13 +17,9 @@ #include "hybrid/node_executor/rts/rts_node_executor.h" #include "hybrid/node_executor/rts/rts_task_factory.h" -#include "common/debug/log.h" #include "common/ge/ge_util.h" -#include "common/types.h" -#include "graph/common/omg_util.h" #include "graph/utils/tensor_utils.h" #include "hybrid/model/hybrid_model.h" -#include "runtime/rt.h" namespace ge { namespace hybrid { @@ -33,6 +29,7 @@ REGISTER_RTS_TASK_CREATOR(IDENTITY, IdentityNodeTask); REGISTER_RTS_TASK_CREATOR(IDENTITYN, IdentityNNodeTask); REGISTER_RTS_TASK_CREATOR(READVARIABLEOP, ReadVariableOpNodeTask); REGISTER_RTS_TASK_CREATOR(PROFILINGTRAININGTRACE, ProfilingTraceNodeTask); +REGISTER_RTS_TASK_CREATOR(MEMCPYASYNC, IdentityNodeTask); Status IdentityNodeTask::DoCopyTensor(TaskContext &context, int index) { auto input_desc = context.MutableInputDesc(index); @@ -133,8 +130,7 @@ Status ProfilingTraceNodeTask::ExecuteAsync(TaskContext &context, std::function< Status RtsNodeExecutor::LoadTask(const HybridModel &model, const NodePtr &node, shared_ptr &task) const { GE_CHECK_NOTNULL(node); GELOGD("[%s] Load for local task.", node->GetName().c_str()); - std::string node_type; - GE_CHK_STATUS_RET(GetOriginalType(node, node_type), "Get original type failed."); + const std::string node_type = NodeUtils::GetNodeType(node); RtsNodeTaskPtr rts_task = RtsTaskFactory::GetInstance().Create(node_type); if (rts_task == nullptr) { GELOGE(UNSUPPORTED, "[%s] Unsupported RTS op type: %s", node->GetName().c_str(), node_type.c_str()); diff --git a/ge/hybrid/node_executor/rts/rts_node_task.cc b/ge/hybrid/node_executor/rts/rts_node_task.cc index 104196ee..aec6804d 100644 --- a/ge/hybrid/node_executor/rts/rts_node_task.cc +++ b/ge/hybrid/node_executor/rts/rts_node_task.cc @@ -43,7 +43,6 @@ namespace hybrid { REGISTER_RTS_TASK_CREATOR(STREAMACTIVE, StreamActiveNodeTask); REGISTER_RTS_TASK_CREATOR(STREAMSWITCH, StreamSwitchNodeTask); REGISTER_RTS_TASK_CREATOR(STREAMMERGE, StreamMergeNodeTask); -REGISTER_RTS_TASK_CREATOR(MEMCPYASYNC, MemcpyAsyncNodeTask); REGISTER_RTS_TASK_CREATOR(ENTER, PassThroughNodeTask); REGISTER_RTS_TASK_CREATOR(REFENTER, PassThroughNodeTask); @@ -168,34 +167,6 @@ Status StreamMergeNodeTask::ExecuteAsync(TaskContext &task_context, std::functio return SUCCESS; } -Status MemcpyAsyncNodeTask::ExecuteAsync(TaskContext &task_context, std::function done_callback) { - GELOGD("[%s] Start to execute.", task_context.GetNodeName()); - auto input_desc = task_context.MutableInputDesc(0); - GE_CHECK_NOTNULL(input_desc); - int64_t copy_size = 0; - GE_CHK_GRAPH_STATUS_RET(TensorUtils::GetTensorSizeInBytes(*input_desc, copy_size)); - // copy_size would not be negative since GetTensorSizeInBytes returned successfully. - if (copy_size > 0) { - const auto in_v = task_context.MutableInput(0); - const auto out_v = task_context.MutableOutput(0); - GE_CHECK_NOTNULL(in_v); - GE_CHECK_NOTNULL(out_v); - GELOGD("[%s] input size: %zu, output size: %zu, copy size: %ld", task_context.GetNodeName(), - in_v->GetSize(), out_v->GetSize(), copy_size); - GE_CHK_RT_RET(rtMemcpyAsync(out_v->MutableData(), out_v->GetSize(), in_v->GetData(), copy_size, - RT_MEMCPY_DEVICE_TO_DEVICE, task_context.GetStream())); - } else { - GELOGW("[%s] invalid copy size: %ld", task_context.GetNodeName(), copy_size); - } - - if (done_callback) { - GE_CHK_STATUS_RET(task_context.RegisterCallback(done_callback)); - } - - GELOGD("[%s] Done executing successfully.", task_context.GetNodeName()); - return SUCCESS; -} - Status PassThroughNodeTask::ExecuteAsync(TaskContext &task_context, std::function done_callback) { GELOGD("[%s] Start to execute.", task_context.GetNodeName()); const auto in_x = task_context.GetInput(0); // x diff --git a/ge/hybrid/node_executor/rts/rts_node_task.h b/ge/hybrid/node_executor/rts/rts_node_task.h index d7d63eb5..e18f9a8f 100644 --- a/ge/hybrid/node_executor/rts/rts_node_task.h +++ b/ge/hybrid/node_executor/rts/rts_node_task.h @@ -60,11 +60,6 @@ class StreamMergeNodeTask : public RtsNodeTask { Status ExecuteAsync(TaskContext &task_context, std::function done_callback) override; }; -class MemcpyAsyncNodeTask : public RtsNodeTask { - public: - Status ExecuteAsync(TaskContext &task_context, std::function done_callback) override; -}; - class PassThroughNodeTask : public RtsNodeTask { public: Status ExecuteAsync(TaskContext &task_context, std::function done_callback) override; diff --git a/ge/hybrid/node_executor/task_context.cc b/ge/hybrid/node_executor/task_context.cc index fe580c1e..bb4340b7 100644 --- a/ge/hybrid/node_executor/task_context.cc +++ b/ge/hybrid/node_executor/task_context.cc @@ -458,10 +458,6 @@ Status TaskContext::PropagateOutputs() { subgraph_context_->all_inputs_[input_offset].SetName( node_item_->NodeName() + "_in_" + std::to_string(dst_input_idx)); } - - auto dst_node_state = subgraph_context_->GetOrCreateNodeState(dst_node_item); - GE_CHECK_NOTNULL(dst_node_state); - dst_node_state->SavePersistTensor(dst_input_idx, *tensor); } } (void)guard; @@ -493,6 +489,7 @@ void TaskContext::ReleaseInputsAndOutputs() { void TaskContext::ReleaseInput(int index) { auto input_tensor = MutableInput(index); if (input_tensor != nullptr) { + node_state_->SavePersistTensor(index, *input_tensor); input_tensor->Destroy(); GELOGD("[%s] Tensor of input[%d] released", GetNodeName(), index); } diff --git a/tests/depends/mmpa/src/mmpa_stub.cc b/tests/depends/mmpa/src/mmpa_stub.cc index aae8de9f..b0f1fb87 100644 --- a/tests/depends/mmpa/src/mmpa_stub.cc +++ b/tests/depends/mmpa/src/mmpa_stub.cc @@ -345,6 +345,10 @@ INT32 mmIsDir(const CHAR *fileName) INT32 mmGetEnv(const CHAR *name, CHAR *value, UINT32 len) { + const char *env = getenv(name); + if (env != nullptr) { + strcpy(value, env); + } return 0; } diff --git a/tests/ut/ge/CMakeLists.txt b/tests/ut/ge/CMakeLists.txt index c5790d59..d55b8861 100755 --- a/tests/ut/ge/CMakeLists.txt +++ b/tests/ut/ge/CMakeLists.txt @@ -726,7 +726,6 @@ set(PASS_TEST_FILES "graph/passes/memcpy_addr_async_unittest.cc" "graph/passes/hccl_continuous_pass_unittest.cc" "graph/passes/hccl_memcpy_pass_unittest.cc" - ) set(KERNEL_TEST_FILES @@ -859,7 +858,6 @@ set(HYBRID_TEST_FILES "hybrid/executor/hybrid_model_async_executor_unittest.cc" "hybrid/executor/hybrid_model_pipeline_executor_unittest.cc" "hybrid/node_executor/aicore/aicore_task_compiler_unittest.cc" - ) set(OTHERS_TEST_FILES @@ -887,6 +885,7 @@ add_library(ge_ut_graph STATIC target_compile_definitions(ge_ut_graph PRIVATE google=ascend_private + FMK_SUPPORT_DUMP ) target_compile_options(ge_ut_graph PRIVATE diff --git a/tests/ut/ge/graph/build/logical_stream_allocator_unittest.cc b/tests/ut/ge/graph/build/logical_stream_allocator_unittest.cc index 218bfd0d..352984fa 100644 --- a/tests/ut/ge/graph/build/logical_stream_allocator_unittest.cc +++ b/tests/ut/ge/graph/build/logical_stream_allocator_unittest.cc @@ -349,7 +349,7 @@ class UtestLogicalStreamAllocator : public testing::Test { /// B --> C(AllReduce) --- D /// / /// stream id: 0 A - /// \ + /// \. /// E --> F(AllReduce) --- G /// stream id: 2 2 2 /// @@ -599,7 +599,7 @@ TEST_F(UtestLogicalStreamAllocator, test_label_not_reusable2) { /// case of multi-output, then unuse stream /// sub1 -/// / | \ +/// / | \. /// sub2 sub3 sub4 TEST_F(UtestLogicalStreamAllocator, test_multiOut_new_stream) { SubGraphInfoPtr data = CreateDataSubgraph(); @@ -624,7 +624,7 @@ TEST_F(UtestLogicalStreamAllocator, test_multiOut_new_stream) { /// if paralle id 1, then use stream /// sub1 -/// / | | \ +/// / | | \. /// sub2 sub3 sub4 sub5 TEST_F(UtestLogicalStreamAllocator, test_parallel_one) { SubGraphInfoPtr data = CreateDataSubgraph(); @@ -653,7 +653,7 @@ TEST_F(UtestLogicalStreamAllocator, test_parallel_one) { /// if the param of engine independent is true, then set independent stream /// sub1 -/// / | | \ +/// / | | \. /// sub2 sub3 sub4 sub5 TEST_F(UtestLogicalStreamAllocator, test_independent) { SubGraphInfoPtr data = CreateDataSubgraph(); @@ -692,7 +692,7 @@ TEST_F(UtestLogicalStreamAllocator, test_independent) { /// set stream based on stream label, and then based on independent /// sub1 -/// / | | \ +/// / | | \. /// sub2 sub3 sub4 sub5 TEST_F(UtestLogicalStreamAllocator, test_independent_switch_label) { SubGraphInfoPtr data = CreateDataSubgraph(); diff --git a/tests/ut/ge/graph/build/stream_allocator_unittest.cc b/tests/ut/ge/graph/build/stream_allocator_unittest.cc index 019e75d1..4ae871af 100644 --- a/tests/ut/ge/graph/build/stream_allocator_unittest.cc +++ b/tests/ut/ge/graph/build/stream_allocator_unittest.cc @@ -36,7 +36,7 @@ class UtestStreamAllocator : public testing::Test { /// /// A - /// / \ + /// / \. /// B C /// | | /// D 400 diff --git a/tests/ut/ge/graph/load/model_manager_unittest.cc b/tests/ut/ge/graph/load/model_manager_unittest.cc index a3545b33..d9e4eabd 100644 --- a/tests/ut/ge/graph/load/model_manager_unittest.cc +++ b/tests/ut/ge/graph/load/model_manager_unittest.cc @@ -438,4 +438,22 @@ TEST_F(UtestModelManagerModelManager, test_data_input_tensor) { auto ret = mm.DataInputTensor(model_id,inputs); EXPECT_EQ(PARAM_INVALID, ret); // HybridDavinciModel::impl_ is null. } + +TEST_F(UtestModelManagerModelManager, test_launch_kernel_cust_aicpu) { + ModelManager mm; + + // cust_aicpu_so_ is empty. + EXPECT_EQ(mm.LaunchKernelCustAicpuSo("empty_cust_aicpu"), SUCCESS); + + // deleteCustOp after Launch will deleted. + uintptr_t resource_id = 1; // for rtCtxGetCurrent stub + std::vector kernel_bin(256); + auto &cust_resource_001 = mm.cust_aicpu_so_[resource_id]; + auto tbe_kernel = std::shared_ptr(new OpKernelBin("deleteCustOp", std::move(kernel_bin))); + auto &cust_opkernel_001 = cust_resource_001["deleteCustOp"] = tbe_kernel; + + EXPECT_FALSE(mm.cust_aicpu_so_.empty()); + EXPECT_EQ(mm.LaunchKernelCustAicpuSo("deleteCustOp"), SUCCESS); + EXPECT_TRUE(mm.cust_aicpu_so_.empty()); +} } // namespace ge diff --git a/tests/ut/ge/graph/passes/assert_pass_unittest.cc b/tests/ut/ge/graph/passes/assert_pass_unittest.cc index 4aa133d3..9247681c 100644 --- a/tests/ut/ge/graph/passes/assert_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/assert_pass_unittest.cc @@ -55,7 +55,7 @@ class UtestGraphPassesAssertPass : public Test { }; /// D E -/// | \ | \ +/// | \ | \. /// F C G /// : | : /// H A I @@ -134,8 +134,8 @@ TEST_F(UtestGraphPassesAssertPass, assert_pass_test2) { EXPECT_EQ(graph->FindNode("D"), nullptr); } -/// E F -/// | \ | \ +/// E F +/// | \ | \. /// H C -> D G /// \ | : /// A I diff --git a/tests/ut/ge/graph/passes/base_pass_unittest.cc b/tests/ut/ge/graph/passes/base_pass_unittest.cc index 9bba5d77..c687e07f 100644 --- a/tests/ut/ge/graph/passes/base_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/base_pass_unittest.cc @@ -130,7 +130,7 @@ class UTESTGraphPassesBasePass : public testing::Test { /// reshape1 /// | /// add1 -/// / \ +/// / \. /// | | /// data1 const1 ComputeGraphPtr BuildGraph1() { @@ -148,9 +148,9 @@ ComputeGraphPtr BuildGraph1() { } /// sum1 -/// / \ -/// / \ -/// / \ +/// / \. +/// / \. +/// / \. /// reshape1 addn1 /// | c | /// add1 <--- shape1 @@ -217,7 +217,7 @@ void CheckIterOrder(UtestTestPass *pass, std::vector &loop, vector &cond) { /******************************************************************************* - * Exit Identify - * \ / \. - * \ / \. - * Switch Add - * / | | - * / | | - * / | | - * LoopCond | | - * \ | | - * \ | | - * \ | | - * Less | | - * \ | NextIteration - * \ | | - * \ | | - * Merge <---------| - * | - * | - * Enter + * | + * +--------------------- Merge ----------------------+ + * / | + * / | + * / | + * / | + * Exit Identify | + * \ / \. | + * \ / \. | + * Switch Add Add + * / | | | + * / | | | + * / | | | + * LoopCond | | | + * \ | | | + * \ | | | + * \ | | | + * Less | | | + * \ | NextIteration | + * \ | | | + * \ | | | + * Merge <---------| | + * | | + * | | + * Enter | + * \ | + * \ | + * Switch Switch + * | | + * +-----------------Equal----------------------+ + * | ******************************************************************************/ - auto data1 = CreateNode(*graph, "data", DATA, 1, 1); + auto data1 = CreateNode(*graph, "data1", DATA, 1, 1); + auto data2 = CreateNode(*graph, "data2", DATA, 1, 1); + + auto equal1 = CreateNode(*graph, "equal1", EQUAL, 2, 1); + auto switch1 = CreateNode(*graph, "switch1", SWITCH, 2, 2); + auto switch2 = CreateNode(*graph, "switch2", SWITCH, 2, 2); + auto enter1 = CreateNode(*graph, "enter", ENTER, 1, 1); - auto merge1 = CreateNode(*graph, "merge", MERGE, 2, 2); - auto less1 = CreateNode(*graph, "less", LESS, 2, 1); + auto merge1 = CreateNode(*graph, "merge1", MERGE, 2, 2); + auto less1 = CreateNode(*graph, "less1", LESS, 2, 1); auto loop1 = CreateNode(*graph, "loopcond", LOOPCOND, 1, 1); - auto switch1 = CreateNode(*graph, "switch", SWITCH, 2, 2); + auto switch3 = CreateNode(*graph, "switch3", SWITCH, 2, 2); auto ident1 = CreateNode(*graph, "identity", IDENTITY, 1, 1); - auto add1 = CreateNode(*graph, "add", ADD, 2, 1); + auto add1 = CreateNode(*graph, "add1", ADD, 2, 1); auto next1 = CreateNode(*graph, "next", NEXTITERATION, 1, 1); auto exit1 = CreateNode(*graph, "exit", EXIT, 1, 1); - auto value0 = CreateNode(*graph, "const", CONSTANT, 0, 1); - auto value1 = CreateNode(*graph, "const", CONSTANT, 0, 1); + auto value1 = CreateNode(*graph, "const1", CONSTANT, 0, 1); + + auto value2 = CreateNode(*graph, "const2", CONSTANT, 0, 1); + auto add2 = CreateNode(*graph, "add2", ADD, 2, 1); + auto merge2 = CreateNode(*graph, "merge2", MERGE, 2, 2); auto output1 = CreateNode(*graph, "net_output", NETOUTPUT, 1, 1); - GraphUtils::AddEdge(data1->GetOutDataAnchor(0), enter1->GetInDataAnchor(0)); + GraphUtils::AddEdge(data1->GetOutDataAnchor(0), equal1->GetInDataAnchor(0)); + GraphUtils::AddEdge(data2->GetOutDataAnchor(0), equal1->GetInDataAnchor(1)); + GraphUtils::AddEdge(data1->GetOutDataAnchor(0), switch1->GetInDataAnchor(0)); + GraphUtils::AddEdge(data2->GetOutDataAnchor(0), switch2->GetInDataAnchor(0)); + GraphUtils::AddEdge(equal1->GetOutDataAnchor(0), switch1->GetInDataAnchor(1)); + GraphUtils::AddEdge(equal1->GetOutDataAnchor(0), switch2->GetInDataAnchor(1)); + cond.emplace_back(switch1); + cond.emplace_back(switch2); + + GraphUtils::AddEdge(switch1->GetOutDataAnchor(0), enter1->GetInDataAnchor(0)); // false GraphUtils::AddEdge(enter1->GetOutDataAnchor(0), merge1->GetInDataAnchor(0)); GraphUtils::AddEdge(merge1->GetOutDataAnchor(0), less1->GetInDataAnchor(0)); GraphUtils::AddEdge(value1->GetOutDataAnchor(0), less1->GetInDataAnchor(1)); GraphUtils::AddEdge(less1->GetOutDataAnchor(0), loop1->GetInDataAnchor(0)); - GraphUtils::AddEdge(loop1->GetOutDataAnchor(0), switch1->GetInDataAnchor(0)); - GraphUtils::AddEdge(merge1->GetOutDataAnchor(0), switch1->GetInDataAnchor(1)); + GraphUtils::AddEdge(loop1->GetOutDataAnchor(0), switch3->GetInDataAnchor(0)); + GraphUtils::AddEdge(merge1->GetOutDataAnchor(0), switch3->GetInDataAnchor(1)); + loop.emplace_back(merge1); - GraphUtils::AddEdge(switch1->GetOutDataAnchor(0), exit1->GetInDataAnchor(0)); - GraphUtils::AddEdge(switch1->GetOutDataAnchor(1), ident1->GetInDataAnchor(0)); + GraphUtils::AddEdge(switch3->GetOutDataAnchor(0), exit1->GetInDataAnchor(0)); // false + GraphUtils::AddEdge(switch3->GetOutDataAnchor(1), ident1->GetInDataAnchor(0)); // true + loop.emplace_back(switch3); GraphUtils::AddEdge(ident1->GetOutDataAnchor(0), add1->GetInDataAnchor(0)); GraphUtils::AddEdge(value1->GetOutDataAnchor(0), add1->GetInDataAnchor(1)); GraphUtils::AddEdge(add1->GetOutDataAnchor(0), next1->GetInDataAnchor(0)); - GraphUtils::AddEdge(next1->GetOutDataAnchor(0), merge1->GetInDataAnchor(1)); - GraphUtils::AddEdge(exit1->GetOutDataAnchor(0), output1->GetInDataAnchor(0)); - merge = merge1; + GraphUtils::AddEdge(switch2->GetOutDataAnchor(1), add2->GetInDataAnchor(1)); // true + GraphUtils::AddEdge(value2->GetOutDataAnchor(0), add2->GetInDataAnchor(0)); + + GraphUtils::AddEdge(exit1->GetOutDataAnchor(0), merge2->GetInDataAnchor(0)); + GraphUtils::AddEdge(add2->GetOutDataAnchor(0), merge2->GetInDataAnchor(1)); + GraphUtils::AddEdge(merge2->GetOutDataAnchor(0), output1->GetInDataAnchor(0)); + + cond.emplace_back(merge2); + merge = merge2; } static void CreateCondGraph(ComputeGraphPtr &graph, NodePtr &merge) { @@ -197,12 +235,24 @@ static void CreateCondGraph(ComputeGraphPtr &graph, NodePtr &merge) { TEST_F(UtestMarkForceUnknownForCondPass, skip_while_loop_merge) { auto graph = std::make_shared("test_graph"); NodePtr merge; - CreateLoopGraph(graph, merge); - - AttrUtils::SetBool(merge->GetOpDesc(), ATTR_NAME_FORCE_UNKNOWN_SHAPE, true); + vector loop; + vector cond; + CreateLoopGraph(graph, merge, loop, cond); MarkForceUnknownForCondPass mark_force_unknown_pass; EXPECT_EQ(mark_force_unknown_pass.Run(graph), SUCCESS); // skip LoopCond + + EXPECT_EQ(loop.size(), 2); + for (const auto &node : loop) { + EXPECT_FALSE(node->GetOpDesc()->HasAttr(ATTR_NAME_CONTROL_FLOW_GROUP)); + } + + EXPECT_EQ(cond.size(), 3); + for (const auto &node : cond) { + int64_t group_index = -1; + EXPECT_TRUE(AttrUtils::GetInt(node->GetOpDesc(), ATTR_NAME_CONTROL_FLOW_GROUP, group_index)); + EXPECT_EQ(group_index, merge->GetOpDesc()->GetId()); + } } TEST_F(UtestMarkForceUnknownForCondPass, skip_known_shape_merge) { diff --git a/tests/ut/ge/graph/passes/merge_pass_unittest.cc b/tests/ut/ge/graph/passes/merge_pass_unittest.cc index 75fdb21b..f8f0afea 100644 --- a/tests/ut/ge/graph/passes/merge_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/merge_pass_unittest.cc @@ -110,8 +110,8 @@ TEST_F(UtestGraphPassesMergePass, multiple_inputs) { } /// Merge -/// | \ -/// | \ +/// | \. +/// | \. /// Op1 Op2 Merge2 /// \ | | /// \ | Op3 @@ -137,10 +137,10 @@ TEST_F(UtestGraphPassesMergePass, empty_input_cut_branch_meet_net_output_with_da } /// Merge -/// | \ -/// | \ +/// | \. +/// | \. /// Op1 Op2 Merge2 -/// \ | | \ +/// \ | | \. /// \ | Op3 /// \ | : /// NetOutput @@ -165,8 +165,8 @@ TEST_F(UtestGraphPassesMergePass, empty_input_cut_branch_meet_net_output_with_co TEST_F(UtestGraphPassesMergePass, empty_input_cut_branch) { /// Merge - /// | \ - /// | \ + /// | \. + /// | \. /// Op1 Op2 Merge2 /// \ | | /// \ | Op3 @@ -210,7 +210,7 @@ TEST_F(UtestGraphPassesMergePass, empty_input_cut_branch) { /// Op1 Op2 Merge2 /// \ | /// \ Op3 - /// \ + /// \. /// Merge3 ret = pass_.Run(merge_node2); @@ -224,7 +224,7 @@ TEST_F(UtestGraphPassesMergePass, single_non_const_input) { /// Op1 /// | /// Merge - /// / \ + /// / \. /// Op2 Op3 auto merge_node = NewNode("Merge", MERGE, 1, 2); auto node1 = NewNode("Op1", RELU, 1, 1); @@ -253,7 +253,7 @@ TEST_F(UtestGraphPassesMergePass, single_const_input) { /// Const /// | /// Merge Pass Const - /// / \ ===> / \ + /// / \ ===> / \. /// Op1 Op2 Op1 Op2 auto merge_node = NewNode("Merge", MERGE, 1, 2); auto const_node = NewNode("Const", CONSTANT, 1, 1); @@ -284,7 +284,7 @@ TEST_F(UtestGraphPassesMergePass, single_const_input_value_index_two_out_nodes) /// / | ===> / \(control anchor) /// Op1 | \ Op1 Constant /// Op2 Op3 | - /// / \ + /// / \. /// Op2 Op3 auto merge_node = NewNode("Merge", MERGE, 1, 2); auto const_node = NewNode("Const", CONSTANT, 1, 1); @@ -329,7 +329,7 @@ TEST_F(UtestGraphPassesMergePass, single_const_input_value_index_two_out_nodes1) /// / | ===> / \(control anchor) /// Op1 | \ Op1 Constant /// Op2 Op3 | - /// / \ + /// / \. /// Op2 Op3 auto merge_node = NewNode("Merge", MERGE, 1, 2); auto const_node = NewNode("Const", CONSTANT, 1, 1); @@ -357,7 +357,7 @@ TEST_F(UtestGraphPassesMergePass, const_with_control_input) { /// C /// | /// Merge - /// / \ + /// / \. /// Op1 Op2 auto switch_node = NewNode("Switch", SWITCH, 1, 2); auto identity_node = NewNode("Identity", SWITCH, 1, 1); @@ -381,7 +381,7 @@ TEST_F(UtestGraphPassesMergePass, const_with_control_input) { /// . /// . /// C - /// / \ + /// / \. /// Op1 Op2 auto ret = pass_.Run(merge_node); EXPECT_EQ(ret, SUCCESS); diff --git a/tests/ut/ge/graph/passes/parallel_group_pass_unittest.cc b/tests/ut/ge/graph/passes/parallel_group_pass_unittest.cc index 588eac9a..a6c3ff6a 100644 --- a/tests/ut/ge/graph/passes/parallel_group_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/parallel_group_pass_unittest.cc @@ -67,11 +67,11 @@ class UtestGraphPassesParallelGgroupPass : public testing::Test { void BuildDefaultGraph() { /// input - /// \ + /// \. /// sqrt pred /// \ / /// cast - /// / \ + /// / \. /// switch_t switch_f /// | | /// F T @@ -119,13 +119,13 @@ class UtestGraphPassesParallelGgroupPass : public testing::Test { void BuildDefaultGraph1() { /// input - /// \ + /// \. /// sqrt pred /// \ / /// Switch /// | | /// ----F T---- - /// \ | / \ + /// \ | / \. /// \ Merge1 Merge2 /// \_________| input_node_ = NewNode("input", RELU, 0, 1); @@ -165,14 +165,14 @@ class UtestGraphPassesParallelGgroupPass : public testing::Test { void BuildDefaultGraph2() { /// input input1 - /// \ \ + /// \ \. /// sqrt pred sqrt1 pred1 /// \ / \ / /// Switch Switch1 /// | | _______| /// | | / /// ____F T____ - /// \ | / \ + /// \ | / \. /// \ Merge1 Merge2 /// \__________| input_node_ = NewNode("input", RELU, 0, 2); diff --git a/tests/ut/ge/graph/passes/reshape_recovery_pass_unittest.cc b/tests/ut/ge/graph/passes/reshape_recovery_pass_unittest.cc index 3be11452..f941645e 100644 --- a/tests/ut/ge/graph/passes/reshape_recovery_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/reshape_recovery_pass_unittest.cc @@ -31,9 +31,9 @@ class UtestReshapeRecoveryPass : public testing::Test { namespace { /// netoutput1 -/// | \ -///transdata1 \ -/// | \ +/// | \. +///transdata1 \. +/// | \. /// | transdata2 /// | / /// var1 const1 diff --git a/tests/ut/ge/graph/passes/reshape_remove_pass_unittest.cc b/tests/ut/ge/graph/passes/reshape_remove_pass_unittest.cc index 351e96d7..ca0cac86 100644 --- a/tests/ut/ge/graph/passes/reshape_remove_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/reshape_remove_pass_unittest.cc @@ -35,7 +35,7 @@ namespace { /// transdata1 /// | /// reshape1 -/// | \ +/// | \. /// var1 const1 ut::GraphBuilder Graph1Builder() { ut::GraphBuilder builder = ut::GraphBuilder("g1"); @@ -55,11 +55,11 @@ ut::GraphBuilder Graph1Builder() { } /// netoutput1 -/// | \ -///transdata1 \ -/// | \ +/// | \. +///transdata1 \. +/// | \. /// reshape1 reshape2 -/// | \ / \ +/// | \ / \. /// var1 const1 var2 ut::GraphBuilder Graph2Builder() { ut::GraphBuilder builder = ut::GraphBuilder("g2"); @@ -83,9 +83,9 @@ ut::GraphBuilder Graph2Builder() { } /// netoutput1 -/// | \ -///transdata1 \ -/// | \ +/// | \. +///transdata1 \. +/// | \. /// reshape1 transdata2 /// | \ / /// var1 const1 diff --git a/tests/ut/ge/graph/passes/resource_pair_control_pass_unittest.cc b/tests/ut/ge/graph/passes/resource_pair_control_pass_unittest.cc index 6d12a49d..8cdfd0c7 100644 --- a/tests/ut/ge/graph/passes/resource_pair_control_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/resource_pair_control_pass_unittest.cc @@ -34,7 +34,7 @@ class UtestResourcePairControlPass : public testing::Test { namespace { /// netoutput1 -/// | \ +/// | \. /// StackPush StackPop /// | | /// var1 const1 diff --git a/tests/ut/ge/graph/passes/switch_logic_remove_pass_unittest.cc b/tests/ut/ge/graph/passes/switch_logic_remove_pass_unittest.cc index dcad318c..22734047 100644 --- a/tests/ut/ge/graph/passes/switch_logic_remove_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/switch_logic_remove_pass_unittest.cc @@ -63,9 +63,9 @@ ComputeGraphPtr BuildGraph1() { /// netoutput1 /// | /// merge1 -/// / \ +/// / \. /// / add1 -/// / F| \ +/// / F| \. /// addn1 swtich2 var3 /// \F T/ | /// switch1 | @@ -101,9 +101,9 @@ ComputeGraphPtr BuildGraph2() { /// add1 /// / \T /// var3 swtich2 -/// T/ \ -/// switch1 \ -/// / \ \ +/// T/ \. +/// switch1 \. +/// / \ \. /// var1 var2 var4 ComputeGraphPtr BuildGraph3() { auto builder = ut::GraphBuilder("g3"); @@ -129,7 +129,7 @@ ComputeGraphPtr BuildGraph3() { /// netoutput1 /// | /// merge1 -/// / \ +/// / \. /// add1 addn1 /// / \T F/ /// var3 swtich2 diff --git a/tests/ut/ge/graph/passes/trans_op_breadth_fusion_pass_unittest.cc b/tests/ut/ge/graph/passes/trans_op_breadth_fusion_pass_unittest.cc index dbb163e1..d05bd695 100644 --- a/tests/ut/ge/graph/passes/trans_op_breadth_fusion_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/trans_op_breadth_fusion_pass_unittest.cc @@ -402,7 +402,7 @@ TEST_F(UtestGraphPassesTransOpBreadthFusionPass, test_multi_anchor_case) { } /// ----> netoutput1 -/// / | \ +/// / | \. /// transdata1 transdata2 transdata3 /// \ / | /// var1-------------- @@ -432,7 +432,7 @@ static ComputeGraphPtr BuildGraph1() { } /// ---------> netoutput1 -/// / | \ +/// / | \. /// transdata1 transdata2(l1) transdata3(l1) /// \ / | /// var1------------------ diff --git a/tests/ut/ge/graph/passes/trans_op_depth_fusion_pass_unittest.cc b/tests/ut/ge/graph/passes/trans_op_depth_fusion_pass_unittest.cc index a9ea41ea..dbac3246 100644 --- a/tests/ut/ge/graph/passes/trans_op_depth_fusion_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/trans_op_depth_fusion_pass_unittest.cc @@ -456,19 +456,19 @@ TEST_F(UtestGraphPassesTransOpDepthFusionPass, test_transop_with_multi_out_edge) /// -->transpose1 -->transpose3-->sinh2 /// | \ / /// | -->transpose2 - /// | \ + /// | \. /// / -->cast3-->cast4-->sinh3 /// / /// / -->transpose4-->transpose5-->sinh4 /// / / /// Node4D-->Cast1-->Cast2-->Cast5 -->reshape2-->sinh5 - /// \ \ + /// \ \. /// \ -->sinh6 - /// \ + /// \. /// \ -->transpose6-->transpose7-->sinh9 /// \ / /// -->reshape-->cast6-->cast7-->sinh8 - /// \ + /// \. /// -->sinh7 /// after optimized graph @@ -479,15 +479,15 @@ TEST_F(UtestGraphPassesTransOpDepthFusionPass, test_transop_with_multi_out_edge) /// / /-->transpose3-->sinh2 /// -->Cast1 /// / \-->sinh7 - /// / \ + /// / \. /// / -->sinh9 /// Node4D /// \ -->sinh4 /// \ / /// -->Cast5-->sinh5 - /// \ \ + /// \ \. /// \ -->sinh6 - /// \ + /// \. /// -->Cast7-->sinh8 ge::ComputeGraphPtr graph = std::make_shared("test"); diff --git a/tests/ut/ge/graph/passes/transop_nearby_allreduce_fusion_pass_unittest.cc b/tests/ut/ge/graph/passes/transop_nearby_allreduce_fusion_pass_unittest.cc index 1220b35e..9c6d8276 100644 --- a/tests/ut/ge/graph/passes/transop_nearby_allreduce_fusion_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/transop_nearby_allreduce_fusion_pass_unittest.cc @@ -180,7 +180,7 @@ ComputeGraphPtr GetGraph7(size_t symmetric_transdata_num, size_t asymmetric_tran /// TransData TransData ... MatMul ... /// \ | / / / /// HcomAllReduce - /// / | \ \ \ + /// / | \ \ \. /// TransData TransData ... RealDiv ... ComputeGraphPtr graph = std::make_shared("test"); NodePtr allreduce = @@ -340,7 +340,7 @@ TEST(UtestTransopNearbyAllreduceFusionPass, test7_all_reduce_with_multiple_trans /// TransData TransData ... MatMul ... /// \ | / / / /// HcomAllReduce - /// / | \ \ \ + /// / | \ \ \. /// TransData TransData ... RealDiv ... size_t symmetric_transdata_num = 20; size_t asymmetric_transdata_num = 20; diff --git a/tests/ut/ge/graph/passes/variable_op_pass_unittest.cc b/tests/ut/ge/graph/passes/variable_op_pass_unittest.cc index f1ea7a27..655867a7 100644 --- a/tests/ut/ge/graph/passes/variable_op_pass_unittest.cc +++ b/tests/ut/ge/graph/passes/variable_op_pass_unittest.cc @@ -66,7 +66,7 @@ namespace { /// transdata2 /// | /// assign1 -/// / \ +/// / \. /// transdata1 | /// | | /// var1 const1 diff --git a/tests/ut/ge/graph/variable_accelerate_ctrl_unittest.cc b/tests/ut/ge/graph/variable_accelerate_ctrl_unittest.cc index 37b4bda7..bf350b6c 100644 --- a/tests/ut/ge/graph/variable_accelerate_ctrl_unittest.cc +++ b/tests/ut/ge/graph/variable_accelerate_ctrl_unittest.cc @@ -35,8 +35,8 @@ namespace { /// shapeNo1 /// | /// addnYes1 -/// / \ -/// / \ +/// / \. +/// / \. /// const1 const2 ComputeGraphPtr BuildGraph1() { @@ -57,9 +57,9 @@ ComputeGraphPtr BuildGraph1() { /// /// netoutput1 -/// / \ \ -/// add1 assign1 \ -/// / \ / \ \ +/// / \ \. +/// add1 assign1 \. +/// / \ / \ \. /// var1 var2 const1 var3 ComputeGraphPtr BuildGraph2() {