OpenGL & Vulkan

[gobrosse]

I have hit a scenario where combining indexing into an array of samplers in a dynamic and non-uniform fashion (allowed via the EXT_descriptor_indexing extension) in combination with a fragment shader making use of the discard statement will systematically cause an application crash due to a lost device on AMD GCN graphics hardware on Windows.

I believe this is due to a driver bug and have taken steps to isolate the problem, I do not have the technical expertise to claim this is absolutely a driver bug and not some wrong-doing on my part, but I have great confidence my program is following the Vulkan specification to the letter and my usage is valid.

I have managed to isolate the problem and can cause it to happen 100% of the time. Here is a github repo containing the binaries and exact shader configuration to trigger this issue.

I've included more precisions inside the readme of that repo

GitHub - Hugobros3/potential-amd-vk-driver-crash

I am available for any precision or follow-up questions about the problem

[dipak]

Hi gobrosse,

Thank you for reporting the problem. I'm moving this thread to our Vulkan forum here: OpenGL & Vulkan.

Also you have been whitelisted now.

Thanks.

[xhuang]

Hello,  would you please provide more info for us to investigate the problem:

1. Driver version? Are you using the latest release version? If not, could you try the latest version ? We had a fix for issues related to 'discard' statement in recent driver releases.

2. Graphics Card Model

3. OS version,etc.

[gobrosse]

Hi, my setup information was provided inside the readme of the linked github repo, along a lot more details than my post contains ( re posting it all here seemed a bit much ). Sorry to not have made it more obvious, but I am indeed using the latest drivers, here's the relevant quote (it's at the end of the readme, in the "further informations" section:

I am using Vulkan 1.1.70 on the latest Adrenalin drivers (18.12.1.1) with a Saphire RX 580 4GB on Windows 8.1, running on a a Ryzen 1700 at stock clocks.

I'll add I'm using the Windows 7 variant of the drivers, but I have reproduced the problem on a 2400G Windows 10 machine, along with a few people on various graphics-programming related communities

[xhuang]

Thank you for your input, we will look into this problem. One more question, is this a regression problem, ie some previous driver release works well with the same application? If you do have the regression information, please let us know. Thanks.

[gobrosse]

The 2400G system I also tested this on is still on 18.8.1, and the problem there seem identical. I'll ask the people who run my reproduction code what driver versions they're on.

[gobrosse]

So I'm pretty sure my bug report had nothing to do with this looking at the timing, but it looks like today's driver update ( 18.12.2 ) did the trick and solved the problem for the time being. Thanks about that !

[xhuang]

Great!

Thank you for your input to help us improve our software and products.