gpu_timeline(python code)

update mindspore/profiler/parser/integrator.py.

update mindspore/profiler/profiling.py.

mindspore/ccsrc/profiler/device/gpu/gpu_profiling.cc

@@ -406,8 +406,10 @@ void GPUProfiler::OpDataProducerBegin(const std::string op_name, void *stream) {
     CHECK_CUDA_RET_WITH_ERROR(cudaEventRecord(op_event_start_, (CUstream)stream_),
                               "cudaEventRecord op event start failed");
     op_host_time_start_ = GetHostTimeStamp();
+    op_cupti_time_start_ = GetCUPTITimeStamp();
   } else {
     op_host_time_start_ = GetHostTimeStamp();
+    op_cupti_time_start_ = GetCUPTITimeStamp();
   }
   SetRunTimeData(op_name, stream);
 }
@@ -431,7 +433,7 @@ void GPUProfiler::OpDataProducerEnd() {
   }
   MS_LOG(DEBUG) << "Host Time Elapsed(us)," << op_name_ << "," << op_time_elapsed;
   SetRunTimeData(op_name_, op_time_elapsed);
-  SetRunTimeData(op_name_, op_host_time_start_, op_time_elapsed);
+  SetRunTimeData(op_name_, op_cupti_time_start_, op_time_elapsed);
 }
 
 void GPUProfiler::StopCUPTI() {

mindspore/ccsrc/profiler/device/gpu/gpu_profiling.h

@@ -170,6 +170,7 @@ class GPUProfiler {
   cudaEvent_t op_event_stop_;
   uint64_t op_host_time_start_;
   uint64_t op_host_time_stop_;
+  uint64_t op_cupti_time_start_;
   std::string profile_data_path_;
 };
 }  // namespace gpu