Crash when using Validation Layers and Vulkan Decode Sample

[mbechard]

Environment:

• OS: Windows 11
• GPU and driver version: RTX Pro 6000 Blackwell
• SDK or header version if building from repo: 1.4.335.0
• Options enabled (synchronization, best practices, etc.): Validation

Describe the Issue

Run the vk-video-dec-test from
https://github.com/nvpro-samples/vk_video_samples
(note you'll need to bring in mbechard/vk_video_samples@4e9f7e8 as of today since the head of main fails to start due to an incorrect error check). Likely that'll be merged/fixed soon though.
With the options:
-i Jelly.mov -v -vv
Using:
https://github.com/user-attachments/assets/700479bc-a791-4af0-a31f-f61a2984fd85

You'll quickly get a crash. The crash sometimes occurs in the Nvidia driver, which I think is caused by the validation layers mangling some of the parameters passed through to the driver.

 	nvoglv64.dll!00007fff64a82d62()	Unknown
 	VkLayer_khronos_validation.dll!vvl::dispatch::Device::QueueSubmit2KHR(VkQueue_T * queue, unsigned int submitCount, const VkSubmitInfo2 * pSubmits, VkFence_T * fence) Line 4772	C++
 	VkLayer_khronos_validation.dll!vulkan_layer_chassis::QueueSubmit2KHR(VkQueue_T * queue, unsigned int submitCount, const VkSubmitInfo2 * pSubmits, VkFence_T * fence) Line 16414	C++
>	vk-video-dec-test.exe!VulkanDeviceContext::MultiThreadedQueueSubmit(const VulkanDeviceContext::QueueFamilySubmitType submitType, const int queueIndex, unsigned int submitCount, const VkSubmitInfo2 * pSubmits, VkFence_T * fence, const char * submissionName, unsigned __int64 decodeEncodeOrder, unsigned __int64 displayInputOrder) Line 193	C++
 	vk-video-dec-test.exe!VkVideoDecoder::DecodePictureWithParameters(VkParserPerFrameDecodeParameters * pCurrFrameDecParams, VkParserDecodePictureInfo * pDecodePictureInfo) Line 1257	C++
 	vk-video-dec-test.exe!NvVulkanDecoder::VulkanVideoParser::DecodePicture(VkParserPictureData * pd, vkPicBuffBase * __formal, VkParserDecodePictureInfo * pDecodePictureInfo) Line 2640	C++
 	vk-video-dec-test.exe!NvVulkanDecoder::VulkanVideoParser::DecodePicture(VkParserPictureData * pd) Line 830	C++
 	nvidia-vkvideo-parser.dll!VulkanVideoDecoder::end_of_picture() Line 513	C++
 	nvidia-vkvideo-parser.dll!VulkanVideoDecoder::nal_unit() Line 364	C++

Other times I get call crashes such as:

 	vcruntime140.dll!_CxxThrowException(void * pExceptionObject, const _s__ThrowInfo * pThrowInfo) Line 80	C++
 	VkLayer_khronos_validation.dll!std::_Throw_future_error2(const std::future_errc _Ec) Line 144	C++
 	[Inline Frame] VkLayer_khronos_validation.dll!std::_Associated_state<int>::_Set_value_raw(int &&) Line 344	C++
 	[Inline Frame] VkLayer_khronos_validation.dll!std::_Associated_state<int>::_Set_value(int &&) Line 338	C++
 	[Inline Frame] VkLayer_khronos_validation.dll!std::_State_manager<int>::_Set_value(int &&) Line 751	C++
 	VkLayer_khronos_validation.dll!std::promise<void>::set_value() Line 1166	C++
 	[Inline Frame] VkLayer_khronos_validation.dll!vvl::Fence::Retire() Line 122	C++
 	VkLayer_khronos_validation.dll!vvl::Queue::Retire(vvl::QueueSubmission & submission) Line 313	C++
 	VkLayer_khronos_validation.dll!vvl::Queue::ThreadFunc() Line 331	C++
 	[Inline Frame] VkLayer_khronos_validation.dll!std::invoke(void(vvl::Queue::*)() &&) Line 1686	C++
 	VkLayer_khronos_validation.dll!std::thread::_Invoke<std::tuple<void (__cdecl vvl::Queue::*)(void),vvl::Queue *>,0,1>(void * _RawVals) Line 61	C++

No crashes occur if validation isn't enabled (remove -v -vv)

[spencer-lunarg]

@mbechard can you also provide the driver version

[aqnuep]

You'll quickly get a crash. The crash sometimes occurs in the Nvidia driver, which I think is caused by the validation layers mangling some of the parameters passed through to the driver.

The actual validation layer code related to video definitely does not change the contents of the parameters passed through to the driver.

@spencer-lunarg, correct me if I'm wrong, but I believe that's also true in general for the validation layers overall, except for the handle wrapping that the validation layers do (I believe by default).

@mbechard, one thing you could check on your end to lock out the possibility of the validation layer handle wrapping causing the issue is to try disabling it (e.g. using vkconfig by unchecking "Handle Wrapping" under the settings of the VK_LAYER_KHRONOS_validation layer).

But even handle wrapping can, I believe, only go wrong if the app uses APIs that the validation layer is not aware of (outdated validation layers or non-public extensions, for example). @spencer-lunarg, can you confirm?

[spencer-lunarg]

@spencer-lunarg, correct me if I'm wrong, but I believe that's also true in general for the validation layers overall, except for the handle wrapping that the validation layers do (I believe by default).

Yes, outside very specific cases where the app provide garbage pointers in pipeline creation, we never (in normal validation) try to adjust any parameters

handle wrapping causing the issue is to try disabling it

You can use the command line set VK_LAYER_UNIQUE_HANDLES=0 as well to disable it

But even handle wrapping can, I believe, only go wrong if the app uses APIs that the validation layer is not aware of

Honestly looking at this more, this might not be a video specific issue, but almost an issue with our Sync state tracking looking at the stack dump ... I will try and run on my AMD Windows machine (and RADV machine) to reproduce locally

[arno-lunarg]

I tried to repro (Win 11, VS 2022, RTX 3070 TI laptop), I get this output and no crash:

Launching Vulkan Application:
- Executable: C:\Users\a\lunarg\dev\Vulkan-Video-Samples\build\vk_video_decoder\demos\Debug\vk-video-dec-test.exe
- Working Directory: C:\Users\a\lunarg\dev\Vulkan-Video-Samples\build\vk_video_decoder\demos\vk-video-dec\
- Command-line Arguments:
  - -i
  - Jelly.mov
  - -v
  - -vv
- Log file: C:\Users\a\VulkanSDK\vk-video-dec-test.txt

0.3.9

Process terminated

As soon as I remove -v -vv I crash though, immediatly:

Launching Vulkan Application:
- Executable: C:\Users\a\lunarg\dev\Vulkan-Video-Samples\build\vk_video_decoder\demos\Debug\vk-video-dec-test.exe
- Working Directory: C:\Users\a\lunarg\dev\Vulkan-Video-Samples\build\vk_video_decoder\demos\vk-video-dec\
- Command-line Arguments:
  - -i
  - Jelly.mov
- Log file: C:\Users\a\VulkanSDK\vk-video-dec-test.txt

Assertion failed: codecType != VK_VIDEO_CODEC_OPERATION_NONE_KHR, file C:\Users\a\lunarg\dev\Vulkan-Video-Samples\vk_video_decoder\libs\VkDecoderUtils\VideoStreamDemuxer.cpp, line 42

Process terminated

Also I had to manually copy nvidia-vkvideo-parser.dll to the directory where vk-video-dec-test.exe is to get things started

[mbechard]

Thanks for the fast response. I'm on another machine today which is a Quadro RTX 5000 (Turing generation). I tried both drivers 582.16 and 591.59. On multiple repeated runs I'm only able to get the crash in the validation layers, not the one in the driver. The crash occurs 100% of the time. It's very strange you are getting a crash without the layers.
Did you also install the FFmpeg libraries so FFmpeg is enabled for the app? You'll also need to copy those .dll files into the .exe's directory.

I'll give more information about the Blackwell machine later today.

[mbechard]

On the Blackwell GPU the driver version is 591.44.
The GPU driver crash is rarer, happening maybe 1/10 of the times. The validation layer crash happens the other 9/10 times. It never gets through a run without one of the two crashes though.
I'm not able to make the GPU crash occur if I turn off Handle Wrapping. I donk think is conclusive since it's a bit random, but seems like a good hint. I think once the validation layer crash is fixed then it'll be easier to narrow down the one that occurs in the GPU driver.

[spencer-lunarg]

tried with AMD Radeon 780M with the latest build of RADV, it is crashing without validation, turning on validation I'm getting errors (like 00384 and 06415) and looking at the RADV stack trace, seems it is just crashing because of doing thing invalidly

[mbechard]

Ya I agree the sample is full of validation errors, which is a whole other issue (especially since isn't this supposed to be the official sample for the extension?)
However the Validation crash makes it hard for me to fix those errors in my port of the code. Do you have an Nvidia GPU you can try on to try to get the validation crash?

[arno-lunarg]

Ah indeed I missed the "install FFmpeg step"
I now repro the crash, consistently. But not when using -i Jelly.mov -v -vv (and without VkConfig).
To repro I removed -v -vv and used VkConfig to setup validation. Crash can occur with Core validation only, those setting options precisely (info mostly targeted at Spencer if he wants to repro):

Crash stack trace I get

 	ucrtbased.dll!00007ffcecab2fb5()	Unknown
 	ucrtbased.dll!00007ffcecab3153()	Unknown
 	ucrtbased.dll!00007ffcecacae2d()	Unknown
 	ucrtbased.dll!00007ffcecac9db0()	Unknown
 	vcruntime140_1d.dll!00007ffdd95b200d()	Unknown
 	vcruntime140_1d.dll!00007ffdd95b2cba()	Unknown
 	vcruntime140_1d.dll!00007ffdd95b2d4a()	Unknown
 	vcruntime140_1d.dll!00007ffdd95b6d3b()	Unknown
 	ntdll.dll!00007ffde71c62ff()	Unknown
 	ntdll.dll!00007ffde7072327()	Unknown
 	ntdll.dll!00007ffde71c5c3e()	Unknown
 	KernelBase.dll!00007ffde45e782a()	Unknown
 	vcruntime140d.dll!00007ffdd2079362()	Unknown
>	VkLayer_khronos_validation.dll!std::_Throw_future_error2(const std::future_errc _Ec) Line 144	C++
 	VkLayer_khronos_validation.dll!std::_Associated_state<int>::_Set_value_raw(int && _Val, std::unique_lock<std::mutex> * _Lock, bool _At_thread_exit) Line 344	C++
 	VkLayer_khronos_validation.dll!std::_Associated_state<int>::_Set_value(int && _Val, bool _At_thread_exit) Line 338	C++
 	VkLayer_khronos_validation.dll!std::_State_manager<int>::_Set_value(int && _Val, bool _Defer) Line 748	C++
 	VkLayer_khronos_validation.dll!std::promise<void>::set_value() Line 1163	C++
 	VkLayer_khronos_validation.dll!vvl::Fence::Retire() Line 123	C++
 	VkLayer_khronos_validation.dll!vvl::Queue::Retire(vvl::QueueSubmission & submission) Line 313	C++
 	VkLayer_khronos_validation.dll!vvl::Queue::ThreadFunc() Line 328	C++
 	VkLayer_khronos_validation.dll!std::invoke<void (__cdecl vvl::Queue::*)(void),vvl::Queue *>(void(vvl::Queue::*)() && _Obj, vvl::Queue * && _Arg1) Line 1712	C++
 	VkLayer_khronos_validation.dll!std::thread::_Invoke<std::tuple<void (__cdecl vvl::Queue::*)(void),vvl::Queue *>,0,1>(void * _RawVals) Line 61	C++
 	ucrtbased.dll!00007ffcecad3010()	Unknown
 	kernel32.dll!00007ffde5c3e8d7()	Unknown
 	ntdll.dll!00007ffde70ec53c()	Unknown

I confirm that turning off validation in VkConfig, just keeping the -i Jelly.mov argument, I don't crash anymore.

[aqnuep]

@mbechard, please file an issue about this on the Vulkan Video Sample repository in order to make sure that the pending issues get fixed there.