Add Gpu profiler interface

paddle/cuda/include/hl_cuda.h

@@ -335,4 +335,14 @@ extern bool hl_cuda_event_is_ready(hl_event_t event);
  */
 extern void hl_device_synchronize();
 
+/**
+ * @brief   gpu profiler start
+ */
+extern void hl_profiler_start();
+
+/**
+ * @brief   gpu profiler stop
+ */
+extern void hl_profiler_end();
+
 #endif  // HL_CUDA_H_

paddle/cuda/include/stub/hl_cuda_stub.h

@@ -93,4 +93,8 @@ inline bool hl_cuda_event_is_ready(hl_event_t event) { return true; }
 
 inline void hl_device_synchronize() {}
 
+inline void hl_profiler_start() {}
+
+inline void hl_profiler_end() {}
+
 #endif  // HL_CUDA_STUB_H_

paddle/cuda/src/hl_cuda_device.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <string.h>
 #include <unistd.h>
 #include <sys/syscall.h>
+#include <cuda_profiler_api.h>
 #include <mutex>
 #include "hl_cuda.h"
 #include "hl_cuda.ph"
@@ -133,8 +134,9 @@ void* cudart_dso_handle = nullptr;
   __macro(cudaGetLastError)               \
   __macro(cudaFuncSetCacheConfig)         \
   __macro(cudaRuntimeGetVersion)          \
-  __macro(cudaGetErrorString)
-
+  __macro(cudaGetErrorString)             \
+  __macro(cudaProfilerStart)              \
+  __macro(cudaProfilerStop)
 CUDA_ROUTINE_EACH(DYNAMIC_LOAD_CUDART_WRAP)
 
 #undef CUDA_ROUNTINE_EACH
@@ -756,3 +758,10 @@ bool hl_cuda_event_is_ready(hl_event_t event) {
   }
   return true;
 }
+
+void hl_profiler_start() {
+    CHECK_CUDA(dynload::cudaProfilerStart());
+}
+void hl_profiler_end() {
+    CHECK_CUDA(dynload::cudaProfilerStop());
+}