Added input/output kernel dump support based on the config file

mindspore/ccsrc/debug/data_dump/dump_json_parser.cc

@@ -307,11 +307,6 @@ void DumpJsonParser::JudgeDumpEnabled() {
   MS_EXCEPTION_IF_NULL(context);
   if (context->get_param<std::string>(MS_CTX_DEVICE_TARGET) == kGPUDevice) {
     async_dump_enabled_ = false;
-    // GPU not support dump kernel inputs
-    if (input_output_ != kDumpOutputOnly) {
-      MS_LOG(WARNING) << "Data dump only support dump kernel output when device target is GPU";
-      input_output_ = kDumpOutputOnly;
-    }
   }
 
   if (context->get_param<std::string>(MS_CTX_DEVICE_TARGET) == kAscendDevice) {

mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc

@@ -102,11 +102,11 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
                     bool dump_enabled) {
   // check if we should read the kernel data
   bool read_data = false;
+  auto &dump_json_parser = DumpJsonParser::GetInstance();
   std::string kernel_name = kernel->fullname_with_scope();
   if (debugger) {
     debugger->SetCurNode(kernel_name);
     if (dump_enabled) {
-      auto &dump_json_parser = DumpJsonParser::GetInstance();
       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
       if ((dump_mode == 0) || ((dump_mode == 1) && dump_json_parser.NeedDump(kernel_name))) {
@@ -120,6 +120,7 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
     return;
   }
 
+  if (dump_json_parser.InputNeedDump()) {
     // get inputs
     auto input_size = AnfAlgo::GetInputTensorNum(kernel);
     for (size_t j = 0; j < input_size; ++j) {
@@ -140,7 +141,9 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
                       << ", tensor_name:" << input_tensor_name << ", host_format:" << format << ".!";
       }
     }
+  }
 
+  if (dump_json_parser.OutputNeedDump()) {
     // get outputs
     auto output_size = AnfAlgo::GetOutputTensorNum(kernel);
     auto node_name = AnfAlgo::GetCNodeName(kernel);
@@ -165,6 +168,7 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
                       << ", tensor_name:" << tensor_name << ", host_format:" << format << ".!";
       }
     }
+  }
 
   debugger->PostExecuteNode();
 }