ncnn::destroy_gpu_instance() causes Segmentation fault on Jetson Nano

[iwatake2222]

I built ncnn library with Vulkan support for Jetson Nano and created a simple project to test it.
Inference worked. (The result was a little different from CPU, but still correct and working)

However, when I call ncnn::destroy_gpu_instance() at the end of code, Segmentation fault (core dumped) occurs.
Does anyone have the same problem or is it only me?

When I set net.use_vulkan_compute = 0 , error doesn't happen. So, I think it definitely caused by GPU.

This is my code: https://github.com/take-iwiw/NcnnMultiPlatformProject/blob/master/MyNcnnProject/TestVulkan.cpp

The following might be another topic, but might be related.
To create ncnn library, I basically followed the instruction described here https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-nvidia-jetson
When I built ncnn, I got Could NOT find Vulkan (missing: Vulkan_LIBRARY Vulkan_INCLUDE_DIR) error. Also, I got link error (e.g. pipeline.cpp:(.text+0xf28): undefined reference to vkCreatePipelineLayout' ).
So, as an ugry workaround, I modified CMakeLists.txt like the following. (/mnt/usbm/Vulkanloader_demos_1.1.10 is the directory I extracted Vulkan_loader_demos_1.1.100.tar.gz. )

if(NCNN_VULKAN)
#    find_package(Vulkan REQUIRED)
#    include_directories(${Vulkan_INCLUDE_DIR})
    include_directories("/mnt/usbm/Vulkanloader_demos_1.1.100/include")
    link_libraries(/usr/lib/aarch64-linux-gnu/libvulkan.so)
endif()

I don't think this is the reason of the problem, though.
If anyone know smarter way to build ncnn librarh for Jetson Nano, please tell me.
I think Vulkan is pre-installed with JetPack, but for some reasons, find_package(Vulkan REQUIRED) doesn't work...

Thank you

[iwatake2222]

I just add the environment info:

• Jetson Nano (jetson-nano-sd-r32.1-2019-03-18)
  • Copied Vulkan_loader_demos_1.1.100.tar.gz
• ncnn version: d46c598

[nihui]

ncnn::destroy_gpu_instance();

the global instance cleanup should be called after all the other instance destructions

you have to create an inner scope for the Net instance, so that the Net instance destructor is called before ncnn::destroy_gpu_instance();

ncnn::create_gpu_instance();

{
    // do your inference here
}

ncnn::destroy_gpu_instance();

[iwatake2222]

@nihui
Thank you for the quick response.
I fixed my code and the error has gone.

Again, thank you for creating the great library!!

[nihui]

#5234

[nihui]

ded0b78