Change fullname_with_scope in dump and online debugger

mindspore/ccsrc/debug/anf_ir_utils.cc

@@ -46,6 +46,15 @@ using mindspore::tensor::TensorPy;
 
 namespace mindspore {
 
+std::string GetKernelNodeName(const AnfNodePtr &anf_node) {
+  std::string kernel_name = anf_node->fullname_with_scope();
+  if (kernel_name.empty()) {
+    kernel_name = anf_node->ToString();
+  }
+  MS_LOG(DEBUG) << "Full scope kernel name is " << kernel_name << ".";
+  return kernel_name;
+}
+
 // ============================================= MindSpore IR Exporter =============================================
 
 std::string AnfExporter::GetNodeType(const AnfNodePtr &nd) {

mindspore/ccsrc/debug/anf_ir_utils.h

@@ -111,6 +111,8 @@ class AnfExporter {
 
 void ExportIR(const std::string &filename, const FuncGraphPtr &func_graph);
 void ExportIR(const std::string &filename, const std::vector<TaggedGraph> &graphs);
+
+std::string GetKernelNodeName(const AnfNodePtr &anf_node);
 }  // namespace mindspore
 
 #endif  // MINDSPORE_CCSRC_DEBUG_ANF_IR_UTILS_H_

mindspore/ccsrc/debug/data_dump/cpu_e2e_dump.cc

@@ -17,12 +17,13 @@
 #include "debug/data_dump/cpu_e2e_dump.h"
 #include <map>
 #include "backend/session/anf_runtime_algorithm.h"
+#include "debug/anf_ir_utils.h"
 
 namespace mindspore {
 void CPUE2eDump::DumpCNodeData(const CNodePtr &node, uint32_t graph_id) {
   MS_EXCEPTION_IF_NULL(node);
   auto &dump_json_parser = DumpJsonParser::GetInstance();
-  std::string kernel_name = node->fullname_with_scope();
+  std::string kernel_name = GetKernelNodeName(node);
   if (!dump_json_parser.NeedDump(kernel_name)) {
     return;
   }
@@ -40,7 +41,7 @@ void CPUE2eDump::DumpCNodeData(const CNodePtr &node, uint32_t graph_id) {
 
 void CPUE2eDump::DumpCNodeInputs(const CNodePtr &node, const std::string &dump_path) {
   MS_EXCEPTION_IF_NULL(node);
-  std::string kernel_name = node->fullname_with_scope();
+  std::string kernel_name = GetKernelNodeName(node);
   MS_LOG(DEBUG) << "Start e2e dump CNode inputs data: " << kernel_name;
   DumpJsonParser::GetInstance().MatchKernel(kernel_name);
   DumpInputImpl(node, dump_path, &kernel_name);
@@ -48,7 +49,7 @@ void CPUE2eDump::DumpCNodeInputs(const CNodePtr &node, const std::string &dump_p
 
 void CPUE2eDump::DumpCNodeOutputs(const CNodePtr &node, const std::string &dump_path) {
   MS_EXCEPTION_IF_NULL(node);
-  std::string kernel_name = node->fullname_with_scope();
+  std::string kernel_name = GetKernelNodeName(node);
   MS_LOG(DEBUG) << "Start e2e dump CNode outputs data: " << kernel_name;
   DumpJsonParser::GetInstance().MatchKernel(kernel_name);
   DumpOutputImpl(node, dump_path, &kernel_name);
@@ -113,7 +114,7 @@ void CPUE2eDump::DumpSingleAnfNode(const AnfNodePtr &anf_node, const size_t outp
   if (!anf_node->isa<Parameter>() && !anf_node->isa<ValueNode>()) {
     return;
   }
-  std::string node_name = anf_node->fullname_with_scope();
+  std::string node_name = GetKernelNodeName(anf_node);
   std::string dump_name = node_name;
   if (anf_node->isa<ValueNode>()) {
     auto iter = const_map->find(node_name);

mindspore/ccsrc/debug/data_dump/dump_json_parser.cc

@@ -21,6 +21,7 @@
 #include "utils/convert_utils_base.h"
 #include "backend/session/anf_runtime_algorithm.h"
 #include "debug/data_dump/npy_header.h"
+#include "debug/anf_ir_utils.h"
 
 namespace {
 constexpr auto kCommonDumpSettings = "common_dump_settings";
@@ -544,15 +545,15 @@ void DumpJsonParser::UpdateNeedDumpKernels(NotNull<const session::KernelGraph *>
   for (const auto &kernel : kernel_graph->execution_order()) {
     MS_EXCEPTION_IF_NULL(kernel);
     if (AnfAlgo::GetKernelType(kernel) == HCCL_KERNEL &&
-        DumpJsonParser::GetInstance().NeedDump(kernel->fullname_with_scope())) {
+        DumpJsonParser::GetInstance().NeedDump(GetKernelNodeName(kernel))) {
       auto input_size = AnfAlgo::GetInputTensorNum(kernel);
       for (size_t i = 0; i < input_size; ++i) {
         auto input_with_index = AnfAlgo::GetPrevNodeOutput(kernel, i);
         auto input = input_with_index.first;
         if (input->isa<CNode>()) {
-          MS_LOG(INFO) << "[AsyncDump] Match Hccl Node:" << kernel->fullname_with_scope()
-                       << " Input:" << input->fullname_with_scope();
-          update_kernels.try_emplace(input->fullname_with_scope(), 0);
+          MS_LOG(INFO) << "[AsyncDump] Match Hccl Node:" << GetKernelNodeName(kernel)
+                       << " Input:" << GetKernelNodeName(input);
+          update_kernels.try_emplace(GetKernelNodeName(input), 0);
         }
       }
     }

mindspore/ccsrc/debug/data_dump/dump_utils.cc

@@ -20,6 +20,7 @@
 
 #include "common/trans.h"
 #include "utils/ms_context.h"
+#include "debug/anf_ir_utils.h"
 #include "debug/data_dump/dump_json_parser.h"
 #include "backend/session/anf_runtime_algorithm.h"
 #include "runtime/device/kernel_runtime_manager.h"
@@ -62,10 +63,11 @@ void SetConstNodeId(const AnfNodePtr &node, std::map<std::string, size_t> *const
   if (!node->isa<ValueNode>()) {
     return;
   }
-  auto iter = const_map->find(node->fullname_with_scope());
+  std::string node_name = GetKernelNodeName(node);
+  auto iter = const_map->find(node_name);
   if (iter == const_map->end()) {
     auto const_idx = const_map->size() + 1;
-    (*const_map)[node->fullname_with_scope()] = const_idx;
+    (*const_map)[node_name] = const_idx;
   }
 }
 

mindspore/ccsrc/debug/data_dump/e2e_dump.cc

@@ -22,6 +22,7 @@
 
 #include "debug/data_dump/dump_json_parser.h"
 #include "common/trans.h"
+#include "debug/anf_ir_utils.h"
 #include "debug/common.h"
 #include "backend/session/anf_runtime_algorithm.h"
 #include "utils/ms_context.h"
@@ -67,7 +68,7 @@ void E2eDump::DumpOutput(const session::KernelGraph *graph, const std::string &d
   const auto &apply_kernels = graph->execution_order();
   for (const auto &node : apply_kernels) {
     MS_EXCEPTION_IF_NULL(node);
-    std::string kernel_name = node->fullname_with_scope();
+    std::string kernel_name = GetKernelNodeName(node);
     if (!dump_json_parser.NeedDump(kernel_name)) {
       continue;
     }
@@ -83,7 +84,7 @@ void E2eDump::DumpOutputSingleNode(const CNodePtr &node, const std::string &dump
   }
   bool trans_flag = dump_json_parser.trans_flag();
   MS_EXCEPTION_IF_NULL(node);
-  std::string kernel_name = node->fullname_with_scope();
+  std::string kernel_name = GetKernelNodeName(node);
   if (!dump_json_parser.NeedDump(kernel_name)) {
     return;
   }
@@ -115,7 +116,7 @@ void E2eDump::DumpOutputImpl(const CNodePtr &node, bool trans_flag, const std::s
                             std::to_string(stream_id) + '.' + std::to_string(timestamp) + ".output." +
                             std::to_string(j);
     if (IsDeviceTargetGPU()) {
-      DumpGPUMemToFile(file_path, node->fullname_with_scope(), *addr, int_shapes, type, device_type, trans_flag, j,
+      DumpGPUMemToFile(file_path, GetKernelNodeName(node), *addr, int_shapes, type, device_type, trans_flag, j,
                        debugger);
     } else {
       DumpMemToFile(file_path, *addr, int_shapes, type, trans_flag);
@@ -134,7 +135,7 @@ void E2eDump::DumpInput(const session::KernelGraph *graph, const std::string &du
   const auto &apply_kernels = graph->execution_order();
   for (const auto &node : apply_kernels) {
     MS_EXCEPTION_IF_NULL(node);
-    std::string kernel_name = node->fullname_with_scope();
+    std::string kernel_name = GetKernelNodeName(node);
     if (!dump_json_parser.NeedDump(kernel_name)) {
       continue;
     }
@@ -150,7 +151,7 @@ void E2eDump::DumpInputSingleNode(const CNodePtr &node, const std::string &dump_
   }
   bool trans_flag = dump_json_parser.trans_flag();
   MS_EXCEPTION_IF_NULL(node);
-  std::string kernel_name = node->fullname_with_scope();
+  std::string kernel_name = GetKernelNodeName(node);
   if (!dump_json_parser.NeedDump(kernel_name)) {
     return;
   }
@@ -177,11 +178,11 @@ void E2eDump::DumpInputImpl(const CNodePtr &node, bool trans_flag, const std::st
     size_t slot;
     if (IsDeviceTargetGPU()) {
       auto input_kernel = node->input(j + 1);
-      std::string input_kernel_name = input_kernel->fullname_with_scope();
+      std::string input_kernel_name = GetKernelNodeName(input_kernel);
       tensor_name = input_kernel_name;
       slot = 0;
     } else {
-      tensor_name = node->fullname_with_scope();
+      tensor_name = GetKernelNodeName(node);
       slot = j;
     }
     ShapeVector int_shapes;
@@ -210,7 +211,7 @@ void E2eDump::DumpSingleAnfNode(const AnfNodePtr &anf_node, const size_t output_
   if ((!anf_node->isa<Parameter>() && !anf_node->isa<ValueNode>()) || IsValueNode<StringImm>(anf_node)) {
     return;
   }
-  std::string node_name = anf_node->fullname_with_scope();
+  std::string node_name = GetKernelNodeName(anf_node);
   std::string dump_name = node_name;
   if (anf_node->isa<ValueNode>()) {
     auto iter = const_map->find(node_name);
@@ -220,11 +221,6 @@ void E2eDump::DumpSingleAnfNode(const AnfNodePtr &anf_node, const size_t output_
     dump_name = std::string("cst") + std::to_string(iter->second);
   }
 
-  // Some parameter nodes have no name. Take the whole string value as the name when dumpping if it's missing.
-  if (dump_name.empty()) {
-    dump_name = anf_node->ToString();
-  }
-
   if (!dump_json_parser.NeedDump(node_name)) {
     return;
   }

mindspore/ccsrc/debug/debug_services.cc

@@ -26,6 +26,7 @@
 #include <unordered_set>
 #include "pybind11/embed.h"
 #ifdef ONLINE_DBG_MODE
+#include "debug/anf_ir_utils.h"
 #include "backend/session/anf_runtime_algorithm.h"
 #endif
 #include "debug/debugger/tensor_summary.h"
@@ -971,7 +972,7 @@ bool DebugServices::IsWatchPointNodeInput(const std::string &w_name, const CNode
     auto input_size = AnfAlgo::GetInputTensorNum(kernel);
     for (size_t j = 0; j < input_size; ++j) {
       auto input_kernel = kernel->input(j + 1);
-      std::string input_kernel_name = input_kernel->fullname_with_scope();
+      std::string input_kernel_name = GetKernelNodeName(input_kernel);
       auto found = w_name.find_last_of('/');
       if (found != std::string::npos && w_name.substr(found + 1) == input_kernel_name) return true;
     }
@@ -1030,7 +1031,7 @@ std::vector<std::shared_ptr<TensorData>> DebugServices::GetNodeTensor(const CNod
   MS_EXCEPTION_IF_NULL(kernel);
   std::vector<std::shared_ptr<TensorData>> result;
   auto output_size = AnfAlgo::GetOutputTensorNum(kernel);
-  auto kernel_name = kernel->fullname_with_scope();
+  auto kernel_name = GetKernelNodeName(kernel);
   for (size_t j = 0; j < output_size; ++j) {
     auto tensor_name_with_slot = kernel_name + ":" + std::to_string(j);
     auto tensor = tensor_loader_->GetTensor(tensor_name_with_slot);

mindspore/ccsrc/debug/debugger/debugger.cc

@@ -37,6 +37,7 @@
 #include "utils/comm_manager.h"
 #include "runtime/hardware/device_context_manager.h"
 #include "debug/anf_ir_dump.h"
+#include "debug/anf_ir_utils.h"
 #ifdef ENABLE_DEBUGGER
 #include "debug/debugger/proto_exporter.h"
 #else
@@ -599,13 +600,13 @@ void Debugger::CheckDatasetGraph() {
   // print parameter node names
   const auto &params = graph_ptr_->inputs();
   for (const auto &param : params) {
-    MS_LOG(INFO) << "param: " << param->fullname_with_scope();
+    MS_LOG(INFO) << "param: " << GetKernelNodeName(param);
   }
   // check if there is GetNext or InitDataSetQueue node
   const auto &nodes = graph_ptr_->execution_order();
   for (const auto &node : nodes) {
     auto node_name = AnfAlgo::GetCNodeName(node);
-    MS_LOG(INFO) << "node: " << node->fullname_with_scope();
+    MS_LOG(INFO) << "node: " << GetKernelNodeName(node);
     if (node_name == "GetNext" || node_name == "InitDataSetQueue") {
       MS_LOG(INFO) << "Not enabling debugger for graph " << graph_ptr_->graph_id() << ": found dataset graph node "
                    << node_name;
@@ -1294,7 +1295,7 @@ void Debugger::LoadSingleAnfnode(const AnfNodePtr &anf_node, const size_t output
   }
   // for parameters and value nodes, set its execution order to be 0;
   int exec_order = 0;
-  std::string node_name = anf_node->fullname_with_scope();
+  std::string node_name = GetKernelNodeName(anf_node);
   GetFileKernelName(NOT_NULL(&node_name));
   // check if output adde exists, if not, return;
   if (!AnfAlgo::OutputAddrExist(anf_node, output_index)) {
@@ -1367,7 +1368,7 @@ void Debugger::LoadGraphOutputs() {
   int exec_order = 1;
   for (const auto &node : apply_kernels) {
     MS_EXCEPTION_IF_NULL(node);
-    std::string kernel_name = node->fullname_with_scope();
+    std::string kernel_name = GetKernelNodeName(node);
     auto output_size = AnfAlgo::GetOutputTensorNum(node);
     if (partial_memory_) {
       if (!debug_services_->IsWatchPoint(kernel_name, node)) {
@@ -1376,7 +1377,7 @@ void Debugger::LoadGraphOutputs() {
     }
     for (size_t j = 0; j < output_size; ++j) {
       if (!AnfAlgo::OutputAddrExist(node, j)) {
-        MS_LOG(INFO) << "Cannot find output addr for slot " << j << " for " << node->fullname_with_scope();
+        MS_LOG(INFO) << "Cannot find output addr for slot " << j << " for " << kernel_name;
         continue;
       }
       auto addr = AnfAlgo::GetOutputAddr(node, j);

mindspore/ccsrc/debug/debugger/debugger_utils.cc

@@ -19,6 +19,7 @@
 #include <vector>
 #include <memory>
 #include <string>
+#include "debug/anf_ir_utils.h"
 #include "debug/debugger/debugger.h"
 #include "runtime/device/gpu/gpu_device_address.h"
 #include "debug/data_dump/dump_json_parser.h"
@@ -58,7 +59,7 @@ void LoadInputs(const CNodePtr &cnode, const KernelLaunchInfo *launch_info_, uin
   auto input_size = AnfAlgo::GetInputTensorNum(cnode);
   for (size_t j = 0; j < input_size; ++j) {
     auto input_kernel = cnode->input(j + 1);
-    std::string input_kernel_name = input_kernel->fullname_with_scope();
+    std::string input_kernel_name = GetKernelNodeName(input_kernel);
     auto addr = kernel_inputs[j];
     auto type = AnfAlgo::GetOutputInferDataType(input_kernel, PARAMETER_OUTPUT_INDEX);
     // For example, this happens with the Depend op
@@ -84,7 +85,7 @@ void LoadOutputs(const CNodePtr &cnode, const KernelLaunchInfo *launch_info_, ui
   auto kernel_outputs = launch_info_->outputs_;
   auto output_size = AnfAlgo::GetOutputTensorNum(cnode);
   auto node_name = AnfAlgo::GetCNodeName(cnode);
-  std::string kernel_name = cnode->fullname_with_scope();
+  std::string kernel_name = GetKernelNodeName(cnode);
   std::vector<int> real_outputs = CheckRealOutput(node_name, output_size);
 
   for (int j : real_outputs) {
@@ -116,7 +117,7 @@ bool CheckReadData(const CNodePtr &cnode) {
   bool read_data = false;
   auto &dump_json_parser = DumpJsonParser::GetInstance();
   bool dump_enabled = debugger->DumpDataEnabledIteration();
-  std::string kernel_name = cnode->fullname_with_scope();
+  std::string kernel_name = GetKernelNodeName(cnode);
   if (dump_enabled) {
     auto dump_mode = dump_json_parser.dump_mode();
     // dump the node if dump_mode is 0, which means all kernels, or if this kernel is in the kernels list

mindspore/ccsrc/debug/debugger/proto_exporter.cc

@@ -23,6 +23,7 @@
 #include <utility>
 #include <algorithm>
 
+#include "debug/anf_ir_utils.h"
 #include "debug/common.h"
 #include "debug/debugger/debugger.h"
 #include "debug/data_dump/dump_json_parser.h"
@@ -436,8 +437,9 @@ void DebuggerProtoExporter::ExportCNode(const FuncGraphPtr &func_graph, const CN
     node_proto->set_scope(node->scope()->name());
 
     // add full_name for debugger
-    node_proto->set_full_name(node->fullname_with_scope());
-    MS_LOG(INFO) << "full_name: " << node->fullname_with_scope();
+    std::string full_name = GetKernelNodeName(node);
+    node_proto->set_full_name(full_name);
+    MS_LOG(INFO) << "full_name: " << full_name;
 
     std::ostringstream buffer;
     auto traces = mindspore::trace::GetSourceLineList(node);

mindspore/ccsrc/debug/dump_proto.cc

@@ -22,6 +22,7 @@
 #include <utility>
 #include <vector>
 
+#include "debug/anf_ir_utils.h"
 #include "debug/common.h"
 #include "proto/anf_ir.pb.h"
 #include "ir/graph_utils.h"
@@ -468,7 +469,7 @@ void ProtoExporter::ExportCNode(const FuncGraphPtr &func_graph, const CNodePtr &
     GetOpNodeTypeAndAttrs(func_graph, op, node_proto);
     node_proto->set_name(std::to_string(apply_idx));
     node_proto->set_scope(node->scope()->name());
-    node_proto->set_full_name(node->fullname_with_scope());
+    node_proto->set_full_name(GetKernelNodeName(node));
 
     // process OP inputs
     for (size_t i = 1; i < inputs.size(); ++i) {