diff --git a/src/gpu.cpp b/src/gpu.cpp
index f5ce57394..eaec7e40a 100644
--- a/src/gpu.cpp
+++ b/src/gpu.cpp
@@ -1612,6 +1612,11 @@ int create_gpu_instance()
     return 0;
 }
 
+VkInstance get_gpu_instance()
+{
+    return (VkInstance)g_instance;
+}
+
 void destroy_gpu_instance()
 {
     MutexLockGuard lock(g_instance_lock);
diff --git a/src/gpu.h b/src/gpu.h
index 497cba44e..4b6484486 100644
--- a/src/gpu.h
+++ b/src/gpu.h
@@ -28,7 +28,19 @@
 namespace ncnn {
 
 // instance
+
+// Create VkInstance and initialize some objects that need to be calculated by GPU
+// Creates a VkInstance object, Checks the extended attributes supported by the Vulkan instance concerned,
+// Initializes, and creates Vulkan validation layers (if ENABLE_VALIDATION_LAYER is enabled),
+// Iterates over all supported physical devices, etc.
 NCNN_EXPORT int create_gpu_instance();
+
+// Get global VkInstance variable
+// Must be called after create_gpu_instance() and before destroy_gpu_instance()
+NCNN_EXPORT VkInstance get_gpu_instance();
+
+// Destroy VkInstance object and free the memory of the associated object
+// Usually called in the destructor of the main program exit
 NCNN_EXPORT void destroy_gpu_instance();
 
 // instance extension capability